Active learning target detection method and device based on complementary active pseudo label
By constructing an input committee and utilizing complementary active pseudo-labels, the information content of candidate boxes is evaluated and manually labeled, thus solving the problems of label redundancy and uncertainty in active learning object detection, and improving detection efficiency and model applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2023-07-17
- Publication Date
- 2026-07-21
AI Technical Summary
Existing active learning object detection methods suffer from redundancy and low efficiency at the bounding box level, difficulty in assessing uncertainty in classification and localization subtasks, high computational cost, and inability to generalize to general detector models.
We adopt a complementary active pseudo-labeling method. By constructing an input committee, we evaluate the information content of candidate boxes using a preset scoring function, select high-information bounding boxes for manual annotation, and update the model by combining sparsely labeled and fully labeled images during the active learning cycle. By leveraging the synergistic effect of manual annotation and pseudo-label generation, we reduce the dependence on annotation.
It effectively reduces the cost of bounding box annotation, improves the generalization ability and efficiency of the detector, reduces computational complexity, and achieves applicability to general detector models.
Smart Images
Figure CN117079086B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of target detection technology, and in particular to an active learning target detection method and apparatus based on complementary active pseudo-labels. Background Technology
[0002] Pool-based active detection strategies rely on scoring functions to label and rank candidate samples, and can be divided into uncertainty-based methods, diversity-based methods, and hybrid methods.
[0003] Among these approaches, uncertainty-based methods prioritize uncertainty predictions based on posterior probability distributions, specific loss prediction modules, or Gaussian mixture heads. To avoid sampling bias caused by uncertainty in batch-based active learning, another strategy is to promote diversity to obtain a more representative dataset, achieved through core set selection. However, diversity-based methods tend to sample as many data points as possible to cover the data manifold, regardless of density. Therefore, hybrid approaches that balance diversity and uncertainty are proposed. In addition to transferring from classification-oriented active learning methods, related techniques also consider the localization subtask, where uncertainty is estimated through inconsistencies between RPN (Region Proposal Network) candidate boxes and final predictions, or through hybrid density models.
[0004] However, the relevant technologies either impose restrictions on the detector architecture or require certain modifications, thus they cannot be generalized to a universal detector model and need further improvement. Summary of the Invention
[0005] This application is based on the inventor's knowledge and discoveries regarding the following issues:
[0006] Based on active scoring functions in related technologies, multi-model methods can fuse different predictions obtained through repeated training, random forward propagation, different model scales, or repeated detection heads. While effective due to increased diversity, fusion methods have not yet become widespread in detection tasks due to computational burden and difficulties in fusion at the bounding box level. Changing model inputs, such as image flipping and noise interference, is simple but has limited impact. In the result fusion process, unlike classification models that can directly average the posterior distribution, detection-oriented transfer methods in related technologies primarily avoid the aforementioned obstacles through image-level scoring and model-level aggregation. However, when applied to the bounding box level, it requires pairwise label assignment, further increasing computational costs.
[0007] While most research on active detection still evaluates at the image level, i.e., the cost is calculated based on detailed annotations of an entire image, analysis reveals that image-level evaluation is impractical and subject to significant evaluation bias. The unit for calculating annotation cost should be bounding boxes. Furthermore, comparisons of active detection methods in related technologies have been conducted without a unified detector (such as SSD, Faster R-CNN, RetinaNet), without standardized learning settings (such as number of iterations, learning rate), with inconsistent benchmark datasets, and with varying levels of supervision (fully supervised or semi-supervised). The resulting conclusions are unreliable and difficult to generalize.
[0008] In summary, this application provides an active learning target detection method and apparatus based on complementary active pseudo-labels to solve the technical problems in related technologies, such as redundancy and low efficiency of bounding box level labels in active learning target detection, and the difficulty in assessing the uncertainty of the two sub-tasks of classification and localization.
[0009] The first aspect of this application provides an active learning target detection method based on complementary active pseudo-labels, comprising the following steps: obtaining an object detector based on the current data pool, and evaluating the object detector to infer an unlabeled data pool; evaluating the information content of candidate boxes in the unlabeled data pool using a preset scoring function, and selecting targets that meet preset screening conditions to obtain manual annotations based on the evaluation results, and adding the manually annotated images to a sparsely labeled data pool or a fully labeled data pool; and updating the object detector using fully labeled and sparsely labeled images or all images in a hybrid supervision setting at the beginning of the first active sampling process, and evaluating both the sparsely labeled images and the unlabeled images during a preset active learning period until the object detector reaches a preset iteration stop condition, so as to achieve a complementary pseudo-label active strategy by utilizing the synergistic effect between manual annotation mining and pseudo-label generation.
[0010] Optionally, in one embodiment of this application, the step of evaluating the information content of candidate boxes in the unlabeled data pool using a preset scoring function includes: constructing a committee for each prediction instance, wherein, while using the exponential moving average of the object detector as the committee chair to generate bounding box reference predictions, an invariance transformation is introduced at the input, and the committee member predictions are determined based on the posterior divergence estimated from multiple random viewpoints of the input, so as to evaluate the information content of the candidate boxes based on the bounding box reference predictions and the committee member predictions.
[0011] Optionally, in one embodiment of this application, after evaluating the information content of the candidate bounding boxes, the method further includes: selecting candidate bounding boxes whose ranking meets preset conditions, so as to identify objects of interest within the candidate region and provide manual annotation for the objects of interest within the candidate region.
[0012] Optionally, in one embodiment of this application, after the first active sampling process ends, the method further includes: when sparsely labeled or unlabeled images are available, appending randomly sampled sparse or unlabeled images to the batch data to generate pseudo-labels; and using the pseudo-labels to supplement sparse or missing real labels to improve the supervision quality of the images after label completion.
[0013] Optionally, in one embodiment of this application, the complementary active pseudo-label strategy achieved by leveraging the synergy between manual annotation mining and pseudo-label generation further includes: For an image with sparse labels, the label-pseudo-active collaboration effect is expressed as: , in, For sparse real bounding boxes, For the location information of sparse real bounding boxes, For the first j The location information of the sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of the pseudo-label bounding boxes that have been filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first The position information of the pseudo-label bounding boxes filtered by category confidence. for Threshold.
[0014] A second aspect of this application provides an active learning target detection device based on complementary active pseudo-labels, comprising: an evaluation module, configured to obtain an object detector based on a current data pool and evaluate the object detector to infer an unlabeled data pool; an acquisition module, configured to evaluate the information content of candidate boxes in the unlabeled data pool using a preset scoring function, select targets that meet preset screening conditions based on the evaluation results to obtain manual annotations, and add the manually annotated images to a sparsely labeled data pool or a fully labeled data pool; and a construction module, configured to update the object detector using fully labeled and sparsely labeled images or all images in a hybrid supervision setting at the beginning of the first active sampling process, and evaluate both the sparsely labeled images and the unlabeled images during a preset active learning period until the object detector reaches a preset iteration stop condition, so as to achieve a complementary pseudo-label active strategy by utilizing the synergy between manual annotation mining and pseudo-label generation.
[0015] Optionally, in one embodiment of this application, the acquisition module includes: an evaluation unit, configured to construct a committee for each prediction instance, wherein, while using the exponential moving average of the object detector as the committee chair to generate a bounding box reference prediction, an invariant transformation is introduced at the input, and the committee member predictions are determined based on the posterior divergence estimated from multiple random viewpoints of the input, so as to evaluate the information content of the candidate box based on the bounding box reference prediction and the committee member predictions.
[0016] Optionally, in one embodiment of this application, after evaluating the information content of the candidate bounding boxes, the acquisition module further includes: a sorting unit, used to select candidate bounding boxes whose ranking meets preset conditions, so as to identify objects of interest within the candidate region and provide manual annotation for the objects of interest within the candidate region.
[0017] Optionally, in one embodiment of this application, the construction module includes: a generation unit, used to append randomly sampled sparse or unlabeled images to batch data to generate pseudo-labels when sparse labeled or unlabeled images are available; and a supplementation unit, used to supplement sparse or missing real labels using the pseudo-labels to improve the supervision quality of the images after label completion.
[0018] Optionally, in one embodiment of this application, the building module further includes: For an image with sparse labels, the label-pseudo-active collaboration effect is expressed as: , in, For sparse real bounding boxes, For the location information of sparse real bounding boxes, For the first j The location information of the sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of the pseudo-label bounding boxes that have been filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first The position information of the pseudo-label bounding boxes filtered by category confidence. for Threshold.
[0019] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the active learning target detection method based on complementary active pseudo-labels as described in the above embodiments.
[0020] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described active learning target detection method based on complementary active pseudo-labels.
[0021] This application embodiment can obtain an object detector based on the current data pool, and infer an unlabeled data pool through evaluation. Then, based on the unlabeled data pool, a preset scoring function is used to evaluate the information content of candidate boxes, resulting in manual annotations. At the beginning of the first active sampling process, sparsely labeled images are determined. The detector is updated using the labeled images or, in a hybrid supervision setting, all images. This leverages the synergy between manual annotation mining and pseudo-label generation to achieve a complementary active pseudo-label strategy. Active learning reduces the dependence on large-scale bounding box level annotations in image object detection. Specifically, the algorithm automatically selects bounding boxes that are helpful for model training for human annotation based on the characteristics of the detection task and data complexity, using limited annotation costs on valuable targets, thereby reducing annotation difficulty and maximizing annotation benefits. Thus, it solves the technical problems of redundant and inefficient bounding box level labels and the difficulty in evaluating the uncertainties of the classification and localization sub-tasks in active learning object detection in related technologies.
[0022] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0023] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an active learning target detection method based on complementary active pseudo-labels provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the evaluation of an active detection method on the VOC0712 dataset under image-level (left) and bounding box-level (right) settings according to an embodiment of this application. Figure 3 This is a schematic diagram illustrating the principle of an active learning target detection method based on complementary active pseudo-labels according to an embodiment of this application; Figure 4 This is a flowchart of an active learning target detection method based on complementary active pseudo-labels according to an embodiment of this application; Figure 5 This is a schematic diagram illustrating the evaluation of an active detection method in VOC-semi, COCO-sup, and COCO-semi settings according to an embodiment of this application. Figure 6 This is a schematic diagram of the structure of an active learning target detection device based on complementary active pseudo-labels according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation
[0024] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0025] The following describes an active learning object detection method and apparatus based on complementary active pseudo-labels according to embodiments of this application, with reference to the accompanying drawings. Addressing the technical problems mentioned in the background art, such as redundancy and low efficiency of bounding box-level labels in active learning object detection, and the difficulty in assessing the uncertainties of the classification and localization sub-tasks, this application provides an active learning object detection method based on complementary active pseudo-labels. In this method, an object detector is obtained based on the current data pool, and an unlabeled data pool is inferred through evaluation. Then, based on the unlabeled data pool, the information content of candidate boxes is evaluated using a preset scoring function to obtain manual annotations. At the beginning of the first active sampling process, sparsely labeled images are determined. The detector is updated using the labeled images or, in a hybrid supervision setting, all images. This leverages the synergy between manual annotation mining and pseudo-label generation to achieve a complementary pseudo-label active strategy. Active learning reduces the dependence on large-scale bounding box-level annotations in image object detection; that is, the algorithm automatically selects bounding boxes that are helpful for model training for human annotation based on the characteristics of the detection task and the complexity of the data, using limited annotation costs on valuable targets, thereby reducing annotation difficulty and maximizing annotation benefits. This solves the technical problems in related technologies, such as the redundancy and low efficiency of bounding box-level labels in active learning object detection, and the difficulty in assessing the uncertainties of the two sub-tasks of classification and localization.
[0026] Specifically, Figure 1 This is a flowchart illustrating an active learning target detection method based on complementary active pseudo-labels provided in an embodiment of this application.
[0027] like Figure 1 As shown, the active learning target detection method based on complementary active pseudo-labels includes the following steps: In step S101, an object detector is obtained based on the current data pool, and the object detector is used for evaluation to infer the unlabeled data pool.
[0028] In actual execution, at the active learning period t=0, the framework of bounding box-level active detection is initialized, a small portion of the images are randomly sampled and fully labeled with bounding boxes, while most images are still unlabeled. Based on the current data pool, this application embodiment can obtain a general object detector, evaluate it, and use it to infer the unlabeled data pool.
[0029] In step S102, the information content of candidate boxes in the unlabeled data pool is evaluated using a preset scoring function, and targets that meet the preset screening conditions are selected based on the evaluation results to obtain manual annotations. The manually annotated images are then added to either the sparsely labeled data pool or the fully labeled data pool.
[0030] Furthermore, in this embodiment, the information content of candidate boxes can be evaluated using a preset scoring function based on an unlabeled data pool, and targets with high information content can be selected to obtain manual annotations based on the evaluation results. The manually annotated images can then be added to a sparsely labeled data pool or a fully labeled data pool, thereby facilitating the synergy between manual annotation mining and simple pseudo-label generation during subsequent active sampling.
[0031] The preset filtering conditions can be set based on the high and low information content thresholds. The high information content threshold can be set by those skilled in the art according to the actual situation, and no specific restrictions are imposed here.
[0032] Unlike image-level active detection methods in related technologies that use images as the smallest annotation unit, the embodiments of this application can actively select candidate bounding boxes with high information content that rank high, allowing annotators to identify objects of interest within the candidate regions. This allows such candidates to proactively prompt annotators to correct potential bounding boxes, rather than passively discovering, labeling, and examining all instances of every class in an image, thereby helping to reduce the spatial search space and semantic options.
[0033] It should be noted that the preset scoring function can be set by those skilled in the art, and no specific restrictions are imposed here.
[0034] Optionally, in one embodiment of this application, evaluating the information content of candidate boxes in an unlabeled data pool using a preset scoring function includes: constructing a committee for each prediction instance, wherein the exponential moving average of the object detector is used as the committee chair to generate bounding box reference predictions while an invariant transformation is introduced at the input, and the committee member predictions are determined based on the posterior divergence estimated from multiple random viewpoints of the input, so as to evaluate the information content of candidate boxes based on the bounding box reference predictions and the committee member predictions.
[0035] like Figure 2As shown, fusion-based active learning has proven effective for detection in classification and bounding box level evaluation. Model fusion, also known as a committee, involves estimating the divergence between the hypotheses of its members. However, fusion strategies in related techniques primarily rely on the repetition of model parameters, a phenomenon known as model-level diversity, which leads to additional computational costs. Furthermore, effectively aggregating bounding box results is complex. Related techniques first adapt the fusion process through instance-level integration to obtain image-level scores, then perform model-level averaging, which is unsuitable for bounding box level detection. Aggregating multiple sets of bounding box predictions leads to pairwise label assignment problems. Therefore, the embodiments of this application must iterate through every prediction of all other members to construct a committee for each instance.
[0036] To reduce computational costs, similar to introducing diversity at the model level, embodiments of this application can introduce invariant transformations at the input level. Therefore, posterior divergence is estimated across multiple random perspectives of the input, which can be considered as committee members. Introducing rich variations from data augmentation rather than model composition significantly reduces the training burden. To reduce the complexity of label assignment, inspired by consensus formulation, embodiments of this application preserve the detector... EMA (exponential moving average) As the committee chair, this serves as a reference for generating bounding box predictions: , in, This indicates the number of training steps within one cycle. For example... Figure 3 As shown, in the embodiment of this application, under the setting of bounding box-level active detection, the proposed complementary pseudo-label active strategy can perform bounding box-level active detection by leveraging the synergy between manual annotation mining and simple pseudo-label generation during active sampling. Iterative optimization is achieved through active acquisition of input committees and complementary training based on pseudo-active collaboration. Therefore, the chairman model effectively detects the input... x More reliable predictions were generated. At the same time, the detectors have taken on more diverse roles, handling a batch of... M A random enhancement Make a hypothesis Based on the chairperson's prediction as a reference, embodiments of this application can measure the discrepancy between its assumptions and those of all other members to effectively reduce the complexity of label assignment. Figure 3 The workflow for sparsely labeled images is shown only to demonstrate its general applicability.
[0037] It is worth noting that these enhancements are input into the network in batches, and in practice, they are computed in parallel rather than forwarded multiple times. Next, the embodiments of this application will describe in detail the quantification of classification and localization divergences.
[0038] Disagreements in classification tasks: When estimating the potential value of a bounding box for a classification branch, embodiments of this application can prioritize disputed regions in the input space. Specifically, given the bounding candidate boxes predicted by the chairperson... Members predict bounding boxes Assigned to For each reference bounding box, use, but is not limited to, the allocation strategy defined by the detector, such as maximum IoU allocation.
[0039] Given a pair of matching bounding boxes The embodiments of this application can be based on single chairman prediction. and posterior predicted membership distribution Cross-entropy is used to measure classification divergence: .
[0040] Regarding bounding box prediction The disagreement is M Summary of all committee members: , in, Indicates the first m The number of member predictions that are a positive match in each random view. The larger this value, the greater the disagreement among the input committee members regarding the predicted candidates for a given bounding box. This indicates that the current model cannot stably make invariant label predictions under different levels of image perturbation, and therefore human annotation should be obtained.
[0041] The divergence in localization tasks: Although the predicted distribution is directly used as the confidence index, This only reflects the committee's disagreement on the classification task. Considering the multi-task nature of detection, the controversy on the localization task should also be considered. When multiple random perturbations are applied to the input, the embodiments of this application can estimate the perturbation of its bounding box prediction regression results. If the predicted location is severely disturbed due to randomness, then the current model's judgment of the target concept is unreliable and should be aided by manual annotation. Conversely, when the prediction results remain stable under changes in the input, it means that the model is already capable of making reliable judgments and does not require manual assistance.
[0042] Specifically, following the chairperson-member label assignment process, a bounding box reference... quilt M Multiple candidates generated by one member The matched bounding boxes are then aligned by applying their respective inverse transforms to them. It is then fed into the chairman model, and the bounding box recalibrated by the chairman is used to measure the... Disagreement on location: , in, The localization branch represents the chairman model. Thus, the localization task is decomposed into four coordinate-based regression tasks. Indicates the first k The prediction standard deviation of each coordinate is normalized by the average of the bounding box height and width.
[0043] In summary, for bounding box level detection tasks, the scoring function in this application embodiment can be described as follows: .
[0044] Measuring disputes in the classification and localization tasks fully leverages bounding box-level human annotations. Based on a scoring function, the construction of the voting committee incorporates input-side randomness to avoid the computational cost of repeated training, while the construction of the reference further reduces the assignment complexity in the proactive bounding box annotation process. As disputed regions in the input space are effectively identified and labeled, the minimization of generalization error is gradually achieved in subsequent loops.
[0045] Optionally, in one embodiment of this application, after evaluating the information content of the candidate bounding boxes, the method further includes: selecting candidate bounding boxes whose ranking meets preset conditions, so as to identify objects of interest within the candidate region and provide manual annotation for the objects of interest within the candidate region.
[0046] Based on the divergence between classification tasks and localization tasks, embodiments of this application can sort the unlabeled regions of all reference boxes. If the top-ranked boxes meet certain IoU-based criteria during the annotation process, manual annotation is provided for the targets of interest within them.
[0047] It should be noted that the preset conditions can be set by those skilled in the art according to different actual situations, and no specific restrictions are imposed here.
[0048] In step S103, at the start of the first active sampling process, the object detector is updated using fully labeled and sparsely labeled images or all images in a hybrid supervised setting. During the preset active learning period, both sparsely labeled and unlabeled images are evaluated until the object detector reaches the preset iteration stop condition, so as to achieve a complementary pseudo-label active strategy by leveraging the synergy between manual annotation mining and pseudo-label generation.
[0049] As one possible way to achieve this, such as Figure 3 As shown, this application embodiment can propose a complementary pseudo-label active strategy under the setting of bounding box level active detection to leverage the synergy between manual annotation mining and simple pseudo-label generation.
[0050] Furthermore, in embodiments of this application, a sparsely labeled image may appear at the start of the first active sampling process, and be updated in each round of active sampling, for example, from the first active sampling process. Initially, sparsely labeled images became available. Then, the detector is updated using the labeled images or, in a hybrid supervision setting, all images. Use only labeled images ( Update, or use all images in a hybrid supervision setting. Update. In the subsequent active learning cycle, the embodiments of this application can evaluate both sparsely labeled and unlabeled images in each round. Among them, bounding boxes with low coincidence with existing manual labels and ranked high are prompted to be labeled, and the process of "training, evaluation, screening, and obtaining manual labels" is repeated until the object detector reaches the preset iteration stop condition. The preset iteration stop condition is not limited to the detector performance threshold, but can also be the result of the budget being exhausted, which can be set by those skilled in the art according to the actual situation, and no specific limitation is made here.
[0051] Optionally, in one embodiment of this application, after the first active sampling process ends, the method further includes: when sparsely labeled or unlabeled images are available, appending randomly sampled sparse or unlabeled images to the batch data to generate pseudo-labels; and using pseudo-labels to supplement sparse or missing real labels to improve the supervision quality of the images after label completion.
[0052] In actual implementation, after the first active sampling cycle, sparsely labeled images are added to the query pool, which not only increases the amount of information but also introduces noise. More seriously, the settings in this application prioritize challenging targets, which, based on experience, are often small, distant, or occluded, while salient, main objects are more likely to remain unlabeled. Therefore, the lack of labels for high-confidence objects provides incorrect supervisory signals, while suggestions surrounding them are incorrectly classified as difficult negative examples. If not handled properly, the sparse labeling problem will have a significant adverse impact on detection performance.
[0053] Despite the serious problem of missing labels, embodiments of this application have provided manual annotations for targets that the detector cannot interpret, while easier targets can be left to machine intelligence. Embodiments of this application design a complementary pseudo-label generation strategy, in which high-confidence model predictions are retained for self-training, while uncertain predictions are filtered out. The complementarity between active sparse annotation and pseudo-label generation can reduce noise caused by missing labels and mitigate the accumulation of errors from false signals.
[0054] Supervised loss for fully labeled images: from Fully tagged image It is fed into the detector and learned in a fully supervised manner: , in, It is the number of fully labeled images. y These are category labels that are manually labeled. t Indicates the position of the corresponding bounding box. and These represent the loss functions used by the detector for classification and localization, respectively. As stated in Formula EMA, the embodiments of this application retain a temporally smoothed version of the detector, namely the chairman model.
[0055] Pseudo-label generation: When sparsely labeled or unlabeled images are available, batches of data are appended with randomly sampled sparse or unlabeled images, a weakly augmented version. x Processed by the chairman model to generate pseudo-labels The enhanced version They are fed into the detector to improve data diversity.
[0056] Similar to the label acquisition strategy in the embodiments of this application, pseudo-labels used for classification and regression are filtered according to different criteria to ensure accuracy. Specifically, embodiments of this application may apply a high confidence threshold. To obtain a reliable bounding box The data is categorized. For localization tasks, embodiments of this application can apply position perturbations to the pseudo-tags. Refine the chairman model, where predicted volatility is below a threshold. The candidate bounding boxes are preserved to supervise the regression head, and these pseudo-labels are represented as .
[0057] Optionally, in one embodiment of this application, the complementary pseudo-label active strategy, which utilizes the synergy between manual annotation mining and pseudo-label generation, further includes: for an image with sparse labels, the label-pseudo-active synergy is expressed as: , in, For sparse real bounding boxes, For the location information of sparse real bounding boxes, For the first j The location information of a sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of pseudo-label bounding boxes filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first Position information of pseudo-label bounding boxes filtered by category confidence. for Threshold.
[0058] In other words, regarding pseudo-labeling-active collaboration in sparse images, although the low recall of pseudo-labels is well-known, and higher confidence thresholds accumulate false negative errors, this situation can be effectively suppressed in the active sparse training setup of this application embodiment. This is because this application embodiment has already provided the annotation of the most challenging targets for an image with sparse labels. The label-pseudo-active synergy effect is represented as follows: , In this application embodiment, pseudo-labels can be used. Supplementing sparse real labels The bounding boxes corresponding to these pseudo-tags The overlap with the real label is less than Jaccard distance. The same deduplication process is also applied to locate branches, thus obtaining... Therefore, the supervision quality of sparse images is improved after label completion: , in, It represents the number of sparsely labeled images.
[0059] Regarding hybrid supervision of unlabeled images, in pool-based active learning scenarios, unlabeled images are also present during training, which can be utilized to improve performance. For images without any human annotations, the loss function is expressed as: , in, This indicates the number of unlabeled images. and These are the pseudo-labels and bounding boxes for the two sub-tasks after threshold filtering.
[0060] Regarding the overall learning objective, in the case of labeled training, the objective function of this application embodiment is expressed as follows: In this application embodiment, a sample ratio can be used. Control the contribution of sparse data streams. Similarly, the objective function for the hybrid supervision setting is... By combining manual annotation and pseudo-labels, the ComPAS active learning strategy utilizes both human intelligence and model intelligence, freeing object detectors from detailed image-level annotation and significantly reducing labor costs.
[0061] Combination Figures 2 to 5 As shown, an embodiment is used to illustrate the working principle and beneficial effects of the active learning target detection method based on complementary active pseudo-labels of this application.
[0062] like Figure 4 As shown, embodiments of this application may include the following steps: Step S401: Problem Definition. In actual execution, at the active learning period t=0, the framework of bounding box-level active detection is initialized, a small portion of the images are randomly sampled and fully labeled with bounding boxes, while most images remain unlabeled. Based on the current data pool, this application embodiment can obtain a general object detector, evaluate it, and use it to infer the unlabeled data pool.
[0063] Unlike image-level active detection methods in related technologies that use images as the smallest annotation unit, the embodiments of this application can actively select candidate bounding boxes with high information content that rank high, allowing annotators to identify objects of interest within the candidate regions. This allows such candidates to proactively prompt annotators to correct potential bounding boxes, rather than passively discovering, labeling, and examining all instances of every class in an image, thereby helping to save labeling costs while reducing the spatial search space and semantic options.
[0064] As one possible way to achieve this, such as Figure 3 The diagram illustrates the bounding box-level active detection process of this application, which iterates between active learning and supplementary training based on pseudo-label-active collaboration through an input committee. Considering general applicability, Figure 3 The workflow of sparsely labeled images is only shown. In the case of bounding box level active detection, this application proposes a complementary pseudo-label active strategy to leverage the synergy between manual annotation mining and simple pseudo-label generation during active sampling.
[0065] Furthermore, in this embodiment, a sparsely labeled image appears at the start of the first active sampling process and is updated in each round of active sampling, for example, from the first active sampling process. Initially, sparsely labeled images became available. Then, the detector is updated using the labeled images or, in a hybrid supervision setting, all images. Use only labeled images ( Update, or use all images in a hybrid supervision setting. Update. In the subsequent active learning cycle, this embodiment of the application can evaluate both the sparsely labeled and unlabeled images in each round, where bounding boxes with low coincidence with existing human-labeled bounding boxes and high rankings are prompted for annotation. The process of "training, evaluation, filtering, and obtaining human annotations" is repeated until the object detector reaches the preset iteration stopping condition.
[0066] Step S402: Active learning based on the input committee. For example... Figure 2 As shown, fusion-based active learning has proven effective for detection under classification and bounding box level evaluation. Active detection methods were evaluated on the VOC0712 dataset under image-level (left) and bounding box level (right) settings. BoxCnt is an attack method designed in this application embodiment. It can far outperform the comparison method simply by marking the image with the most potential predicted candidate boxes, but it is ineffective in bounding box level evaluation. This shows that image-level evaluation is very biased, while bounding box level evaluation is more objective and realistic. The method sCOMPAS proposed in this application embodiment significantly outperforms the comparison method.
[0067] Model fusion, also known as a committee process, involves estimating the divergence between the hypotheses of its members. However, fusion strategies in related techniques primarily rely on the repetition of model parameters, a phenomenon known as model-level diversity, which incurs additional computational costs. Furthermore, effectively aggregating bounding box results is complex. Related techniques first adapt the fusion process through instance-level integration to obtain image-level scores, then perform model-level averaging, which is unsuitable for bounding box-level detection. Aggregating multiple sets of bounding box predictions leads to pairwise label assignment problems; therefore, the embodiments of this application must iterate through every prediction of all other members to construct a committee for each instance.
[0068] To reduce computational costs, similar to introducing diversity at the model level, embodiments of this application can introduce invariant transformations at the input level. Therefore, posterior divergence is estimated across multiple random perspectives of the input, which can be considered as committee members. Introducing rich variations from data augmentation rather than model composition significantly reduces the training burden. To reduce the complexity of label assignment, inspired by consensus formulation, embodiments of this application preserve the detector... EMA (exponential moving average) As the committee chair, this serves as a reference for generating bounding box predictions: , in, This indicates the number of training steps within one cycle. For example... Figure 3 As shown, the chairman model responds to the input x More reliable predictions were generated. At the same time, the detectors have taken on more diverse roles, handling a batch of... M A random enhancement Make a hypothesis Based on the chairman's predictions as a reference, embodiments of this application can measure the discrepancies between its assumptions and those of all other members to effectively reduce the complexity of label assignment.
[0069] It is worth noting that these enhancements are input into the network in batches, and in practice, they are computed in parallel rather than forwarded multiple times. Next, the embodiments of this application will describe in detail the quantification of classification and localization divergences.
[0070] Disagreements in classification tasks: When estimating the potential value of a bounding box for a classification branch, embodiments of this application can prioritize disputed regions in the input space. Specifically, given the bounding candidate boxes predicted by the chairperson... Members predict bounding boxes Assigned to For each reference bounding box, use, but is not limited to, the allocation strategy defined by the detector, such as maximum IoU allocation.
[0071] Given a pair of matching bounding boxes The embodiments of this application can be based on single chairman prediction. and posterior predicted membership distribution Cross-entropy is used to measure classification divergence: .
[0072] Regarding bounding box prediction The disagreement is M Summary of all committee members: , in, Indicates the first m The number of member predictions that are a positive match in each random view. The larger this value, the greater the disagreement among the input committee members regarding the predicted candidates for a given bounding box. This indicates that the current model cannot stably make invariant label predictions under different levels of image perturbation, and therefore human annotation should be obtained.
[0073] The divergence in localization tasks: Although the predicted distribution is directly used as the confidence index, This only reflects the committee's disagreement on the classification task. Considering the multi-task nature of detection, the controversy on the localization task should also be considered. When multiple random perturbations are applied to the input, the embodiments of this application can estimate the perturbation of its bounding box prediction regression results. If the predicted location is severely disturbed due to randomness, then the current model's judgment of the target concept is unreliable and should be aided by manual annotation. Conversely, when the prediction results remain stable under changes in the input, it means that the model is already capable of making reliable judgments and does not require manual assistance.
[0074] Specifically, following the chairperson-member label assignment process, a bounding box reference... quilt M Multiple candidates generated by one member The matched bounding boxes are then aligned by applying their respective inverse transforms to them. It is then fed into the chairman model, and the bounding box recalibrated by the chairman is used to measure the... Disagreement on position: , in, The localization branch represents the chairman model. Thus, the localization task is decomposed into four coordinate-based regression tasks. Indicates the first k The prediction standard deviation of each coordinate is normalized by the average of the bounding box height and width.
[0075] In summary, for bounding box level detection tasks, the scoring function in this application embodiment can be described as follows: .
[0076] Based on this, the embodiments of this application can sort the unlabeled regions of all reference boxes. If the top-ranked boxes meet certain IoU-based criteria during the annotation process, manual annotation can be provided for the targets of interest in them.
[0077] Measuring disputes in the classification and localization tasks fully leverages bounding box-level human annotations. Based on a scoring function, the construction of the voting committee incorporates input-side randomness to avoid the computational cost of repeated training, while the construction of the reference further reduces the assignment complexity in the proactive bounding box annotation process. As disputed regions in the input space are effectively identified and labeled, the minimization of generalization error is gradually achieved in subsequent loops.
[0078] Step S403: Sparse and Hybrid Supervised Training. In actual execution, after the first active sampling loop, sparsely labeled images are added to the query pool, which not only increases the amount of information but also introduces noise. More seriously, the settings of this embodiment prioritize challenging targets, which, based on experience, are often small, distant, or occluded, while salient, main objects are more likely to be unlabeled. Therefore, the lack of labels for high-confidence objects provides incorrect supervisory signals, while suggestions around them are incorrectly classified as difficult negative examples. If not handled properly, the sparse labeling problem will have a significant adverse impact on detection performance.
[0079] Despite the serious problem of missing labels, embodiments of this application have provided manual annotations for targets that the detector cannot interpret, while easier targets can be left to machine intelligence. Embodiments of this application design a complementary pseudo-label generation strategy, in which high-confidence model predictions are retained for self-training, while uncertain predictions are filtered out. The complementarity between active sparse annotation and pseudo-label generation can reduce noise caused by missing labels and mitigate the accumulation of errors from false signals.
[0080] Supervised loss for fully labeled images: from Fully tagged image It is fed into the detector and learned in a fully supervised manner: , in, It is the number of fully labeled images. y These are category labels that are manually labeled. t Indicates the position of the corresponding bounding box. and These represent the loss functions used by the detector for classification and localization, respectively. As stated in Formula EMA, the embodiments of this application retain a temporally smoothed version of the detector, namely the chairman model.
[0081] Pseudo-label generation: When sparsely labeled or unlabeled images are available, batches of data are appended with randomly sampled sparse or unlabeled images, a weakly augmented version. x Processed by the chairman model to generate pseudo-labels The enhanced version They are fed into the detector to improve data diversity.
[0082] Similar to the label acquisition strategy in the embodiments of this application, pseudo-labels used for classification and regression are filtered according to different criteria to ensure accuracy. Specifically, embodiments of this application may apply a high confidence threshold. To obtain a reliable bounding box The data is categorized. For localization tasks, embodiments of this application can apply position perturbations to the pseudo-tags. Refine the chairman model, where predicted volatility is below a threshold. The candidate bounding boxes are preserved to supervise the regression head, and these pseudo-labels are represented as .
[0083] Pseudo-labeling for sparse images - Active collaboration: Although the low recall of pseudo-labels is well-known, and higher confidence thresholds accumulate false negatives, this can be effectively suppressed in the active sparse training setup of this application embodiment. This is because this application embodiment already provides annotation for the most challenging targets, for an image with sparse labels. The label-pseudo-active synergy effect is represented as follows: , In this application embodiment, pseudo-labels can be used. Supplementing sparse real labels The bounding boxes corresponding to these pseudo-tags The overlap with the real label is less than Jaccard distance. The same deduplication process is also applied to locate branches, thus obtaining... Therefore, the supervision quality of sparse images is improved after label completion: , in, It represents the number of sparsely labeled images.
[0084] Hybrid supervision of unlabeled images: In pool-based active learning scenarios, unlabeled images are also present during training, which can be utilized to improve performance. For images without any human annotations, the loss function is expressed as:
[0085] , in, Indicates the number of unlabeled images. and These are the pseudo-labels and bounding boxes for the two sub-tasks after threshold filtering.
[0086] Overall learning objective: In the case of labeled training, the objective function of this application embodiment is expressed as follows: In this application embodiment, a sample ratio can be used. Control the contribution of sparse data streams. Similarly, the objective function for the hybrid supervision setting is... By combining manual annotation and pseudo-labels, the ComPAS active learning strategy utilizes both human intelligence and model intelligence, freeing object detectors from detailed image-level annotation and significantly reducing labor costs.
[0087] In practical applications, such as Figure 5 As shown, where Figure 5 From left to right, these are the bounding box comparison results for VOC-semi, COCO-sup, and COCO-semi. VOC and COCO are detection datasets, while sup and semi represent labeled supervised training and hybrid supervised training, respectively.
[0088] Based on the embodiments of this application, within a unified codebase, ComPAS consistently outperformed 10 comparable methods across four training settings. Under supervision with only labeled data, it achieved 100% supervised performance on VOC0712 with only 19% of bounding box-level annotations, while on the COCO dataset, it surpassed the second-best method by up to 4.3% mAP. ComPAS also supports training with unlabeled data pools, achieving an 85% label reduction rate that surpasses 90% of the COCO supervised performance.
[0089] The active learning object detection method based on complementary active pseudo-labels proposed in this application can obtain an object detector based on the current data pool, infer an unlabeled data pool through evaluation, and then evaluate the information content of candidate boxes using a preset scoring function based on the unlabeled data pool to obtain manual annotations. At the beginning of the first active sampling process, sparsely labeled images are determined, and the detector is updated using the labeled images or all images in a hybrid supervision setting. This achieves a complementary pseudo-label active strategy by leveraging the synergy between manual annotation mining and pseudo-label generation. Active learning reduces the dependence on large-scale bounding box level annotations in image object detection. That is, the algorithm automatically selects bounding boxes that are helpful for model object classification and localization training based on the characteristics of the detection task and the complexity of the data for human annotation, using limited annotation costs on valuable targets, thereby reducing annotation difficulty and maximizing annotation benefits. Thus, it solves the technical problems of bounding box level label redundancy and low efficiency, and the difficulty in evaluating the uncertainty of the two sub-tasks of classification and localization in active learning object detection in related technologies.
[0090] Next, referring to the accompanying drawings, we describe the active learning target detection device based on complementary active pseudo-labels according to the embodiments of this application.
[0091] Figure 6 This is a block diagram of an active learning target detection device based on complementary active pseudo-labels according to an embodiment of this application.
[0092] like Figure 6As shown, the active learning target detection device 10 based on complementary active pseudo-labels includes: an evaluation module 100, an acquisition module 200, and a construction module 300.
[0093] Specifically, the evaluation module 100 is used to obtain an object detector based on the current data pool and to evaluate the object detector in order to infer the unlabeled data pool.
[0094] The acquisition module 200 is used to evaluate the information content of candidate boxes in the unlabeled data pool using a preset scoring function, select targets that meet the preset screening conditions based on the evaluation results to obtain manual annotations, and add the manually annotated images to the sparsely labeled data pool or the fully labeled data pool.
[0095] The construction module 300 is used to update the object detector at the beginning of the first active sampling process using fully labeled and sparsely labeled images or all images in a hybrid supervised setting, and to evaluate both sparsely labeled and unlabeled images during a preset active learning cycle until the object detector reaches a preset iteration stopping condition, so as to achieve a complementary pseudo-label active strategy by leveraging the synergy between manual annotation mining and pseudo-label generation.
[0096] Optionally, in one embodiment of this application, the acquisition module 200 includes an evaluation unit.
[0097] The evaluation unit is used to construct a committee for each prediction instance. It utilizes the exponential moving average of the object detector as the committee chair to generate bounding box reference predictions while introducing an invariant transformation at the input. The committee member predictions are determined based on the posterior divergence estimated from multiple random viewpoints of the input. The informativeness of the candidate boxes is evaluated based on the bounding box reference predictions and the committee member predictions.
[0098] Optionally, in one embodiment of this application, after evaluating the amount of information in the candidate boxes, the acquisition module 200 further includes a sorting unit.
[0099] The sorting unit is used to select candidate bounding boxes whose ranking meets the preset conditions, so as to identify objects of interest within the candidate region and provide manual annotation for the objects of interest within the candidate region.
[0100] Optionally, in one embodiment of this application, the construction module 300 includes a generation unit and a supplementary unit.
[0101] The generation unit is used to append randomly sampled sparse or unlabeled images to batch data to generate pseudo-labels when sparsely labeled or unlabeled images are available.
[0102] The supplementary unit is used to supplement sparse or missing real labels with pseudo-labels to improve the supervision quality of the image after label completion.
[0103] Optionally, in one embodiment of this application, the construction module 300 further includes: For an image with sparse labels, the label-pseudo-active collaboration effect is expressed as: , in, For sparse real bounding boxes, For the location information of sparse real bounding boxes, For the first j The location information of a sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of pseudo-label bounding boxes filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first Position information of pseudo-label bounding boxes filtered by category confidence. for Threshold.
[0104] It should be noted that the foregoing explanation of the active learning target detection method based on complementary active pseudo-labels also applies to the active learning target detection device based on complementary active pseudo-labels in this embodiment, and will not be repeated here.
[0105] The active learning target detection device based on complementary active pseudo-labels proposed in this application can obtain an object detector from the current data pool, infer an unlabeled data pool through evaluation, and then evaluate the information content of candidate boxes using a preset scoring function based on the unlabeled data pool to obtain human annotations. At the beginning of the first active sampling process, sparsely labeled images are determined, and the detector is updated using the labeled images or all images in a hybrid supervision setting. This reduces the dependence on large-scale bounding box level annotations in image target detection by utilizing active learning. That is, the algorithm automatically selects bounding boxes that are helpful for model target classification and localization training based on the characteristics of the detection task and the complexity of the data for human annotation, using limited annotation costs on valuable targets, thereby reducing annotation difficulty and maximizing annotation benefits. Thus, it solves the technical problems of bounding box level label redundancy and low efficiency, and the difficulty in evaluating the uncertainty of the two sub-tasks of classification and localization in active learning target detection in related technologies.
[0106] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 701, the processor 702, and the computer program stored on the memory 701 and executable on the processor 702.
[0107] When the processor 702 executes the program, it implements the active learning target detection method based on complementary active pseudo-labels provided in the above embodiments.
[0108] Furthermore, electronic devices also include: Communication interface 703 is used for communication between memory 701 and processor 702.
[0109] The memory 701 is used to store computer programs that can run on the processor 702.
[0110] The memory 701 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.
[0111] If the memory 701, processor 702, and communication interface 703 are implemented independently, then the communication interface 703, memory 701, and processor 702 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.
[0112] Optionally, in a specific implementation, if the memory 701, processor 702, and communication interface 703 are integrated on a single chip, then the memory 701, processor 702, and communication interface 703 can communicate with each other through an internal interface.
[0113] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.
[0114] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described active learning target detection method based on complementary active pseudo-labels.
[0115] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0116] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0117] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.
[0118] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.
[0119] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0120] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0121] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0122] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
Claims
1. A method for active learning target detection based on complementary active pseudo-labels, characterized in that, Includes the following steps: An object detector is obtained based on the current data pool, and the object detector is used for evaluation to infer the unlabeled data pool; The information content of candidate boxes in the unlabeled data pool is evaluated using a preset scoring function, and targets that meet the preset screening conditions are selected based on the evaluation results to obtain manual annotations. The manually annotated images are then added to either the sparsely labeled data pool or the fully labeled data pool. as well as At the start of the first active sampling process, the object detector is updated using fully labeled and sparsely labeled images or all images in a hybrid supervised setting. During a preset active learning cycle, both the sparsely labeled and unlabeled images are evaluated until the object detector reaches a preset iteration stop condition, so as to achieve a complementary pseudo-label active strategy by leveraging the synergy between manual annotation mining and pseudo-label generation. The step of evaluating the information content of candidate boxes in the unlabeled data pool using a preset scoring function includes: constructing a committee for each prediction instance, wherein the exponential moving average of the object detector is used as the committee chair to generate bounding box reference predictions, while an invariant transformation is introduced at the input, and the committee member predictions are determined based on the posterior divergence estimated from multiple random viewpoints of the input, so as to evaluate the information content of the candidate boxes based on the bounding box reference predictions and the committee member predictions.
2. The method according to claim 1, characterized in that, After evaluating the information content of the candidate boxes, the following is also included: Candidate bounding boxes that meet the preset ranking criteria are selected to identify objects of interest within the candidate regions, and manual annotations are provided for these objects within the candidate regions.
3. The method according to claim 1, characterized in that, After the first active sampling process is completed, the process also includes: When sparsely labeled or unlabeled images are available, randomly sampled sparse or unlabeled images are appended to the batch data to generate pseudo-labels; The pseudo-labels are used to supplement sparse or missing real labels, thereby improving the supervision quality of the image after label completion.
4. The method according to claim 3, characterized in that, The complementary active pseudo-label strategy, which leverages the synergy between manual annotation mining and pseudo-label generation, also includes: For an image with sparse labels, the label-pseudo-active collaboration effect is expressed as: ， in, For sparse real bounding boxes, For the position information of sparse real bounding boxes, For the first j The location information of the sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of the pseudo-label bounding boxes that have been filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first The position information of the pseudo-label bounding boxes filtered by category confidence. for Threshold.
5. An active learning target detection device based on complementary active pseudo-labels, characterized in that, include: An evaluation module is used to obtain an object detector based on the current data pool and to evaluate the object detector to infer the unlabeled data pool. The acquisition module is used to evaluate the information content of candidate boxes in the unlabeled data pool using a preset scoring function, select targets that meet the preset screening conditions according to the evaluation results, obtain manual annotations, and add the manually annotated images to the sparsely labeled data pool or the fully labeled data pool. as well as The module is used to update the detector at the beginning of the first active sampling process using fully labeled and sparsely labeled images or all images in a hybrid supervised setting, and to evaluate both the sparsely labeled and unlabeled images during a preset active learning period until the object detector reaches a preset iteration stop condition, so as to achieve a complementary pseudo-label active strategy by leveraging the synergy between manual annotation mining and pseudo-label generation. The acquisition module includes an evaluation unit for constructing a committee for each prediction instance. The evaluation unit uses the exponential moving average of the object detector as the committee chair to generate a bounding box reference prediction while introducing an invariant transformation at the input. The evaluation unit determines the committee member predictions based on the posterior divergence estimated from multiple random viewpoints of the input. The evaluation unit assesses the information content of the candidate box based on the bounding box reference prediction and the committee member predictions.
6. The apparatus according to claim 5, characterized in that, After evaluating the information content of the candidate boxes, the acquisition module further includes: The sorting unit is used to select candidate bounding boxes that meet the preset ranking criteria, so as to identify objects of interest within the candidate region and provide manual annotation for the objects of interest within the candidate region.
7. The apparatus according to claim 5, characterized in that, The building module includes: The generation unit is used to append randomly sampled sparse or unlabeled images to batch data to generate pseudo-labels when sparsely labeled or unlabeled images are available. The supplementary unit is used to supplement sparse or missing real labels using the pseudo-labels, so as to improve the supervision quality of the image after label completion.
8. The apparatus according to claim 7, characterized in that, The building module also includes: For an image with sparse labels, the label-pseudo-active collaboration effect is expressed as: ， in, For sparse real bounding boxes, For the position information of sparse real bounding boxes, For the first j The location information of the sparse real bounding box. The category of the pseudo-label bounding box after category confidence filtering. For the first The categories of the pseudo-label bounding boxes that have been filtered by category confidence. Position information of pseudo-label bounding boxes after category confidence filtering. For the first The position information of the pseudo-label bounding boxes filtered by category confidence. for Threshold.
9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the active learning target detection method based on complementary active pseudo-labels as described in any one of claims 1-4.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the active learning target detection method based on complementary active pseudo-labels as described in any one of claims 1-4.