Passive domain adaptive target detection method and system based on dual enhancement
By employing a dual-enhancement framework and a fast-slow teacher collaborative architecture, the problem of pseudo-label quality degradation in passive domain adaptive object detection is solved, achieving a balance between model stability and adaptability, and improving detection accuracy.
Patent Information
- Application Number
- CN202511175944.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2025-12-05
AI Technical Summary
In passive domain adaptive object detection, the mean teacher framework suffers from the problem of continuous degradation of pseudo-label quality, causing the model to get stuck in local optima and making it difficult to balance stability and adaptability.
A dual-enhancement framework is adopted, which uses a bidirectional knowledge transfer mechanism in the back-distillation and forward distillation stages, combined with a fast and slow teacher collaborative architecture, and utilizes external basic model knowledge guidance and internal model collaborative optimization to achieve hierarchical fusion of domain-specific knowledge and general visual priors. A self-learning mechanism for pseudo-label fusion and a periodic parameter reset mechanism are used to suppress pseudo-label quality degradation.
It effectively suppresses the degradation of pseudo-label quality and improves the performance of the model in adaptive object detection in unsupervised domains. It is significantly better than existing methods, achieving a balance between model stability and adaptability, and improving detection accuracy.
Smart Images
Figure CN121074584A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer vision, and particularly relates to a passive domain adaptive target detection method and system based on double enhancement. BACKGROUND
[0002] Under the constraint condition that the source domain data is completely inaccessible, the traditional feature alignment method based on adversarial training is difficult to be directly applied due to the dependence on the source domain data distribution information. The current SF-DAOD research generally adopts the Mean Teacher (MT) framework as the basic architecture: the teacher model parameters are maintained through the Exponential Moving Average (EMA) mechanism, and the relatively stable feature representation is used to generate pseudo-labels, which in turn guide the optimization process of the student model.
[0003] However, the Mean Teacher framework has an essential defect, that is, the continuous degradation of pseudo-label quality. In the closed-loop iterative process of "teacher model generates pseudo-labels → student model learns pseudo-labels → student model updates teacher model parameters", the labeling error in the early stage will be continuously accumulated and amplified through the closed-loop parameter update path, eventually leading the model to fall into a local optimum and lose the improvement potential. SUMMARY
[0004] In order to solve the above problems, the present application proposes a passive domain adaptive target detection method and system based on double enhancement, proposes a double enhancement framework, and innovatively combines the double enhancement mechanism of external basic model knowledge guidance and internal model collaborative optimization to break through the defects of the prior art. For external guidance, a bidirectional distillation mechanism is proposed, which realizes the hierarchical deep fusion of domain-specific knowledge and general visual prior by constructing a bidirectional knowledge interaction channel between the task-specific detector and the general basic model.
[0005] The technical scheme adopted by the present application is as follows: A passive domain adaptive target detection method based on double enhancement, comprising: A reverse distillation phase: a reverse knowledge transfer path from a source domain pre-trained target detector to a general basic model is established; based on the knowledge transfer of pseudo-labels, the task-related knowledge of the target detector is transferred to the general basic model, so that the general basic model adapts to the unlabeled target domain; A forward distillation phase: a cross-modal alignment supervision signal is generated based on the adapted general basic model, and a fast-slow double teacher collaborative architecture with a periodic parameter reset mechanism is used to train the target detector, so that the target detector adapts to the unlabeled target domain.
[0006] Furthermore, the fast and slow dual-teacher collaborative architecture includes a fast teacher model, a slow teacher model, and a student model, wherein the fast teacher model and the slow teacher model are respectively connected to the student model.
[0007] Furthermore, the update frequency of the fast teacher model is higher than the first threshold, and it captures dynamic changes in the distribution of target domain data through agile parameter adjustments, generating more domain-related pseudo-labels; the update frequency of the slow teacher model is lower than the second threshold, serving as a stable knowledge anchor to maintain the robustness of the model's core features; the first threshold is greater than the second threshold.
[0008] Furthermore, the periodic parameter reset mechanism includes: performing a fast teacher model state reset every K iterations, and using the stable parameters of the slow teacher model as a benchmark to prevent noise accumulation.
[0009] Furthermore, both the reverse distillation stage and the forward distillation stage employ a self-learning mechanism based on pseudo-label fusion. In the backdistillation stage, the general base model enhances the supervision of the target detector by integrating autonomously discovered task-aware patterns, evolving into a hybrid detector that integrates open-world knowledge and task expertise. In the forward distillation stage, the target detector, guided by pseudo-labels, autonomously integrates the learned domain features with the inherent generalization of the general base model and task-adaptive knowledge.
[0010] Furthermore, the pseudo-label generation method in the backdistillation stage includes:
[0011] in, This indicates a pseudo-label, NMS indicates non-maximum suppression, and CAT indicates detection result splicing. Indicates threshold Filtered target detector pseudo-labels, Indicates threshold The general basic model for filtering is pseudo-labels.
[0012] Furthermore, the pseudo-label generation method in the forward distillation stage includes:
[0013] in, This indicates a pseudo-label, NMS indicates non-maximum suppression, and CAT indicates detection result splicing. Indicates threshold Filtered fast teacher model pseudo-labels, Indicates threshold After filtering and adaptation, the general basic model pseudo-labels are used.
[0014] Furthermore, in the backdistillation stage, the goal is to adapt the general base model to the unlabeled target domain, and the loss function includes:
[0015] in, It is the original detection loss function of the general base model, used to calculate the prediction results of the general base model. With pseudo-tags The losses between them.
[0016] Furthermore, in the forward distillation stage, the goal is to adapt the object detector to the unlabeled target domain under the guidance of the adapted general base model. The loss function used to optimize the object detector includes:
[0017] in, This is the original detection loss function of the object detector, used to calculate the prediction results of the student model. With pseudo-tags The losses between them.
[0018] A passive domain adaptive target detection system based on dual enhancement includes: The backdistillation module is configured to establish a reverse knowledge transfer path from the pre-trained target detector in the source domain to the general base model; based on pseudo-label knowledge transfer, the task-related knowledge of the target detector is transferred to the general base model, making the general base model adaptable to the unlabeled target domain. The forward distillation module is configured to generate cross-modal aligned supervision signals based on the adapted general base model, and train the target detector through a fast and slow dual-teacher collaborative architecture with a periodic parameter reset mechanism, thereby adapting the target detector to the unlabeled target domain.
[0019] The beneficial effects of this invention are as follows: (1) This invention can solve the error accumulation amplification effect of mean teacher framework in unsupervised adaptive target detection. It proposes a dual enhancement innovation framework that integrates external multimodal basic model guidance and internal fast and slow teacher collaborative optimization. This framework can effectively suppress the problem of pseudo-label quality degradation by constructing a dual enhancement mechanism of internal and external collaboration.
[0020] (2) This invention designs an innovative paradigm of bidirectional knowledge distillation and establishes a closed-loop knowledge interaction channel between task-specific detectors and general basic models. In the back-distillation stage, the task-specific knowledge of the source domain detector is transferred back to the general basic model; in the forward distillation stage, the enhanced general basic model is used to generate cross-modal alignment supervision signals, forming a virtuous cycle of bidirectional knowledge enhancement.
[0021] (3) This invention proposes a fast and slow dual-teacher collaborative architecture with a periodic parameter reset mechanism, which can overcome the theoretical problem that it is difficult to achieve both model stability and adaptability in the traditional mean-teacher framework. This invention innovatively designs a differentiated update strategy: the fast teacher uses high-frequency parameter updates to capture dynamic changes in the target domain, while the slow teacher maintains the stability of core features through ultra-low-frequency updates; the periodic parameter synchronization mechanism achieves a dynamic balance between noise suppression and feature exploration by resetting the fast teacher to the slow teacher's baseline state.
[0022] (4) Experiments conducted on multiple object detection cross-domain transfer datasets show that the present invention achieves significant performance improvement, which is significantly better than the state-of-the-art methods. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of a passive domain adaptive target detection method based on dual enhancement, according to Embodiment 2 of the present invention.
[0024] Figure 2 This is a schematic diagram of a general basic model G-DINO according to Embodiment 2 of the present invention.
[0025] Figure 3 This is a schematic diagram of a fast and slow teacher framework according to Embodiment 2 of the present invention. Detailed Implementation
[0026] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0027] Example 1 This embodiment provides a passive domain adaptive target detection method based on dual enhancement, including: Backdistillation stage: Establish a reverse knowledge transfer path from the pre-trained target detector in the source domain to the general base model; based on pseudo-label knowledge transfer, transfer the task-related knowledge of the target detector to the general base model, improve the general base model's ability to perceive tasks, and enable the general base model to adapt to the unlabeled target domain. Forward distillation stage: Based on the adapted general base model, cross-modal aligned supervision signals are generated, and the target detector is trained through a fast and slow dual-teacher collaborative architecture with a periodic parameter reset mechanism, thereby adapting the target detector to the unlabeled target domain. This stage forms a forward guidance mechanism from the general base model to the target detector, ensuring efficient knowledge transfer and performance optimization from the general base model to the target detector, and strengthening the forward guidance capability from the general base model to the target detector.
[0028] To address the core contradiction of balancing model stability and adaptability in the traditional mean-teacher framework, this embodiment proposes a fast-slow dual-teacher collaborative framework with a periodic parameter reset mechanism. This architecture includes two parallel branches for differentiated updates: The fast teacher branch uses a high update frequency and captures dynamic changes in the distribution of target domain data through agile parameter adjustments, such as sudden environmental disturbances or morphological variations of target objects. The slow teacher branch employs an ultra-low update rate as a stable knowledge anchor to maintain the robustness of the model's core features.
[0029] The periodic parameter reset mechanism refers to resetting the fast teacher branch state every K iterations, using the stable parameters of the slow teacher branch as a benchmark, thereby achieving a dual optimization objective: (1) Error correction function: The reliable parameters of the slow teacher branch suppress the annotation noise accumulated by the fast teacher branch during the rapid update process, and prevent model degradation caused by error propagation; (2) Explore the enhanced function: the reset fast teacher branch can break through the constraints of local optimal solutions and restart the global exploration of the target domain feature space. This periodic restart strategy effectively balances the model convergence speed and generalization ability.
[0030] Example 2 This embodiment is based on embodiment 1: This embodiment provides a passive domain adaptive target detection method based on dual enhancement, including external base model guidance and internal fast and slow teacher adjustment mechanisms. For example... Figure 1 As shown, the base model guidance is achieved through bidirectional distillation, which aims to adapt the base model to a specific task and generate high-quality pseudo-labels; the fast and slow teacher control mechanism is used in the forward distillation stage to perform a robust self-training process, as explained below.
[0031] I. External Basic Model Guidance like Figure 2As shown, the general-purpose basic model G-DINO will be used as an example for illustration. Although G-DINO demonstrates excellent zero-shot capability in cross-domain detection, its deployment in unseen domains is limited by a lack of task-specific knowledge. Crucially, if the original G-DINO is directly used as the teacher model, these limitations will be passed on to the student detector, disrupting the target domain adaptation stage. Therefore, this embodiment proposes a bidirectional distillation paradigm (back-distillation + forward distillation) to overcome the domain adaptation bottleneck by establishing bidirectional knowledge transfer.
[0032] In this process, backdistillation constructs an unconventional knowledge transfer mechanism: despite the more compact structure of the source domain pre-trained detector, it can provide specialized task guidance to the general-purpose base model G-DINO. This stage strategically injects task-specific knowledge from the source detector into G-DINO through pseudo-label-based knowledge transfer, enabling it to evolve into a general-purpose-specific hybrid with both open-world understanding and task-optimized detection capabilities.
[0033] Forward distillation utilizes the adapted G-DINO as a dual-knowledge integrator, combining its task expertise with its inherent cross-modal capabilities for target domain adaptation. The calibrated model, through task knowledge obtained via backdistillation, strategically guides its general knowledge base to achieve task-aware semantic alignment between text and image pairs. This collaborative integration enables the adapted G-DINO to generate high-quality pseudo-labels, thereby improving the supervision quality required by the self-training framework.
[0034] Furthermore, to overcome the imperfections of the static teacher model in bidirectional distillation, this embodiment employs a self-learning mechanism of pseudo-label fusion in both stages. Specifically, in the backdistillation stage, G-DINO enhances the supervision of the source detector by integrating autonomously discovered task-aware patterns, evolving into a hybrid detector that integrates open-world knowledge and task expertise. In the forward distillation stage, the target detector, guided by pseudo-labels, autonomously fuses learned domain features with G-DINO's dual capabilities (inherent generalization + task-adaptive knowledge). It should be noted that backdistillation and forward distillation are executed sequentially rather than cyclically.
[0035] Given G-DINO and the source pre-trained detector, the pseudo-labels generated during the backdistillation stage are as follows:
[0036] in, This indicates a pseudo-label, NMS indicates non-maximum suppression, and CAT indicates detection result splicing. Indicates threshold Filtered source detector pseudo-labels, Indicates threshold Filtered G-DINO pseudo-labels.
[0037] The target detector is trained using a fast-slow teacher mechanism during the forward distillation stage. Given the adapted G-DINO and the initialized target detector, the pseudo-label generation formula is:
[0038] in, Indicates threshold Filtered fast teacher model pseudo-labels, Indicates threshold Filtering adaptation for G-DINO pseudo-labels.
[0039] II. Internal teacher speed adjustment mechanism Existing domain adaptation methods using mean teachers have limitations due to the dual challenges of fixed smoothing coefficients: excessively large coefficients can delay the absorption of new domain knowledge due to overly smoothed parameters, while excessively small coefficients can reduce robustness to training noise due to drastic parameter fluctuations.
[0040] Therefore, this embodiment proposes a fast and slow teacher framework, such as... Figure 3 As shown in the figure, the fast teacher model rapidly absorbs target domain knowledge through a low-smoothness EMA update mechanism, achieving efficient domain adaptation; while the slow teacher model adopts a conservative update strategy with a high-smoothness EMA as a stable reference benchmark.
[0041] The core innovation of this framework lies in reinitializing the parameters of the fast teacher model at key training nodes. This design offers dual advantages: 1) it generates more domain-relevant pseudo-labels through the agile adaptability of the fast teacher model; 2) it prevents noise accumulation by periodically refreshing the fast teacher model using the stable parameters of the slow teacher model. The synergistic effect of these two approaches ensures domain adaptation efficiency while effectively resisting label noise interference.
[0042] In practical implementation, the parameters of the fast teacher model... Use a lower smoothing coefficient Updated parameters for the slow-speed teacher model. Use a higher smoothing coefficient During the middle of training, the parameters of the slow teacher model are synchronized to those of the fast teacher model:
[0043] Where t represents the current training round, and T represents the total training rounds.
[0044] III. Optimization Objectives During the reverse distillation stage, the goal is to adapt G-DINO to the unlabeled target domain. This embodiment employs LoRA (low-rank adaptation) technology for efficient fine-tuning of G-DINO, specifically allowing for setting the rank... and scaling factor The loss function can be expressed as:
[0045] in, This is the original detection loss function of G-DINO, used to calculate the G-DINO prediction results. With pseudo-tags The losses between them.
[0046] In the forward distillation stage, the goal is to adapt the target detector to the unlabeled target domain under the guidance of the adapted G-DINO. The loss function used to optimize the target detector can be expressed as:
[0047] in, This is the original detection loss function of DETR, used to calculate the prediction results of the student model. With pseudo-tags The losses between them.
[0048] IV. Implementation Details Specifically, this embodiment uses Deformable DETR as the base detector, and guides this base detector using the large base model G-DINO. In the backdistillation stage, LoRA technology is employed, and G-DINO is fine-tuned using the AdamW optimizer (initial learning rate 5e-4). The threshold is set to... and During the forward distillation phase, the smoothing coefficients for the fast teacher model and the slow teacher model were set to 0.999 and 0.9996, respectively (0.9999 for the latter in the cross-weather adaptation task). Threshold and The student model was optimized using the Adam optimizer (initial learning rate 2e-4). The mean average precision (mAP) with a threshold of 0.5 was used as the evaluation metric, and mAP was calculated using the detection results from the slow teacher model.
[0049] Experimental validation has been conducted on several typical transfer tasks, and the experimental results are shown in Tables 1 to 3. In the tables, "This method*" indicates that Vanilla G-DINO is used as the teacher model.
[0050] Table 1: Comparison of Experimental Results for the Cityscape → FoggyCityscape Migration Task
[0051] Table 1 shows that backdistillation improved the mAP of G-DINO from 31.3% to 45.7%, proving that task knowledge injection is the key to stimulating the domain adaptation potential of the base model; forward distillation further pushed the target detector performance to 46.8% mAP, surpassing the existing state-of-the-art and DRU of the same architecture; the target detector based on native G-DINO guidance only achieved 41.3% mAP, indicating that SF-DAOD must first inject task knowledge into the base model through a bidirectional transfer mechanism, and then transfer its customized knowledge to the detector.
[0052] Table 2: Comparison of Experimental Results for the Cityscape→BDD100K Migration Task
[0053] Table 2 shows that the native G-DINO achieves a zero-shot performance of 41.4% mAP, surpassing the existing state-of-the-art (SOTA). Its knowledge transfer enables the object detector to achieve a 41.0% mAP. The knowledge transfer adapted by this method boosts the object detector performance to 48.6% mAP, an improvement of 12.0% over the SOTA and 7.6% over the native G-DINO-guided object detector. This demonstrates that although the basic model has strong zero-shot capability, task adaptation is still indispensable (as evidenced by the 7.6% difference).
[0054] Table 3: Comparison of Experimental Results for the Sim10K→Cityscape Transfer Task
[0055] Table 3 shows that in the sunny data domain (only detecting the "car" category), the native G-DINO achieved an AP of 79.3%, and the task-adapted version achieved an AP of 80.3%. The dual-enhancement framework guided by it achieved a peak AP of 74.8%, which is more than 15.0% higher than the existing state-of-the-art methods, verifying the advantages of the basic model and the external guidance mechanism.
[0056] In summary, as shown in Tables 1 to 3, the passive domain adaptive target detection method based on dual enhancement in this embodiment achieves optimal performance.
[0057] Example 3 This embodiment provides a passive domain adaptive target detection system based on dual enhancements, including: The backdistillation module is configured to establish a reverse knowledge transfer path from the pre-trained target detector in the source domain to the general base model; based on pseudo-label knowledge transfer, the task-related knowledge of the target detector is transferred to the general base model, making the general base model adaptable to the unlabeled target domain. The forward distillation module is configured to generate cross-modal aligned supervision signals based on the adapted general base model, and train the target detector through a fast and slow dual-teacher collaborative architecture with a periodic parameter reset mechanism, thereby adapting the target detector to the unlabeled target domain.
[0058] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the passive domain adaptive target detection method based on dual enhancements described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.
[0059] Example 5 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the passive domain adaptive target detection method based on dual enhancements described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content contained in the storage medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.
[0060] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
[0061] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
Claims
1. A dual boosting based passive domain adaptive object detection method, characterized in that, Comprise: a reverse distillation stage: establishing a reverse knowledge transfer path from a source domain pre-trained target detector to a general base model; a pseudo-label based knowledge transfer, transferring task related knowledge of the target detector to the general base model to adapt the general base model to the unlabeled target domain; a forward distillation stage: generating a cross-modal alignment supervision signal based on the adapted general base model, and training the target detector through a fast and slow double teacher collaborative architecture with a periodic parameter reset mechanism, so as to adapt the target detector to the unlabeled target domain.
2. The dual boosting based passive domain adaptive object detection method according to claim 1, wherein, The fast and slow double teacher collaborative architecture comprises a fast teacher model, a slow teacher model and a student model, and the fast teacher model and the slow teacher model are connected to the student model.
3. The dual boosting based passive domain adaptive object detection method according to claim 2, wherein, The update frequency of the fast teacher model is higher than a first threshold value, the dynamic changes in the target domain data distribution are captured through agile parameter adjustment, and more field related pseudo labels are generated; the update frequency of the slow teacher model is lower than a second threshold value, and the slow teacher model serves as a stable knowledge anchor to maintain the robustness of the model core features; the first threshold value is greater than the second threshold value.
4. The dual boosting based passive domain adaptive object detection method according to claim 2, wherein, The periodic parameter reset mechanism comprises: resetting the state of the fast teacher model every K iterations, and taking the stable parameters of the slow teacher model as a benchmark to prevent noise accumulation.
5. The dual boosting based passive domain adaptive object detection method according to claim 1, wherein, In the reverse distillation stage and the forward distillation stage, a self-learning mechanism of pseudo label fusion is adopted: In the reverse distillation stage, the general base model enhances the supervision of the target detector by integrating the task perception mode discovered autonomously, and evolves into a hybrid detector that combines open world knowledge and task expertise; In the forward distillation stage, the target detector autonomously integrates the domain features learned and the inherent generalization and task adaptation knowledge of the general base model guided by the pseudo label.
6. The dual boosting based passive domain adaptive object detection method according to claim 5, wherein, The pseudo label generation method in the reverse distillation stage comprises: wherein, denotes pseudo labels, NMS denotes non-maximum suppression, CAT denotes concatenation of detection results, denotes threshold filtered target detector pseudo labels, denotes threshold filtered generic base model pseudo labels.
7. The dual boosting based passive domain adaptive object detection method according to claim 5, wherein, The pseudo label generation method in the forward distillation stage comprises: wherein, denotes pseudo labels, NMS denotes non-maximum suppression, CAT denotes concatenation of detection results, denotes threshold filtered fast teacher model pseudo labels, denotes threshold filtered adapted general base model pseudo labels.
8. The dual boosting based passive domain adaptive object detection method according to claim 1, wherein, In the reverse distillation stage, the goal is to adapt the general base model to the unlabeled target domain, and the loss function comprises: wherein, is the original detection loss function of the general base model, used to calculate the loss between the prediction result of the general base model and the pseudo label .
9. The dual boosting based passive domain adaptive object detection method according to claim 1, wherein, In the forward distillation stage, the goal is to adapt the target detector to the unlabeled target domain under the guidance of the adapted general base model, and the loss function for optimizing the target detector comprises: wherein, is the original detection loss function of the target detector, used to calculate the prediction result of the student model and the pseudo label between.
10. A dual augmentation based passive domain adaptive target detection system, comprising: Comprise: a reverse distillation module configured to establish a reverse knowledge transfer path from a source domain pre-trained target detector to a general base model; a pseudo-label based knowledge transfer, transferring task related knowledge of the target detector to the general base model to adapt the general base model to the unlabeled target domain; a forward distillation module configured to generate a cross-modal alignment supervision signal based on the adapted general base model, and train the target detector through a fast and slow double teacher collaborative architecture with a periodic parameter reset mechanism, so as to adapt the target detector to the unlabeled target domain.