Method for determining pose of target object in image

By combining the coarse splitter and dense splitter methods, the standardized object model and target object representation are generated, which solves the problem of machine learning model overfitting specific objects, and improves the accuracy of pose detection for unseen objects and the confusion of poses Scenario processing capability.

CN119027428APending Publication Date: 2024-11-26ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410653667.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-26
Filing Date
2024-05-24
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing machine learning models tend to overfit specific objects when detecting target objects in images, resulting in poor generalization capabilities for unseen objects and require partial retraining of the model to improve performance.

Method used

Using a combination method of at least two segmenters, a coarse segmenter is used to initially divide images, and a dense segmenter is used to feature extraction and distinguish target objects. It is fine-tuned through a small number of reference images to generate a standardized object model and target object representation, combined with local reconstruction Perform pose evaluation with the object model and optimize segmenter parameters to improve contrast and similarity.

Benefits of technology

It improves the accuracy of pose detection for unseen objects, reduces dependence on the model, enhances the ability to separate target objects from other objects in messy scenes, and improves the accuracy of the final pose.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027428A_ABST
    Figure CN119027428A_ABST
Patent Text Reader

Abstract

A method for determining a pose of a target object in an input image includes providing a set of reference images containing the target object and annotated with respective poses of the target object; inputting the reference image to at least one divider configured to distinguish portions and / or features of the images belonging to different objects; determining a normative object model from the output of the at least one divider, the normative object model aggregating information from the reference image in a common coordinate system; determining a representation of the target object in an output space of the at least one divider according to the output of the at least one divider in combination with the annotation of the reference image; inputting an input image to the at least one divider and determining which portions of the resulting output of the at least one divider belong to the target object based on a comparison of the resulting output of the at least one divider to the representation of the target object; and combining the parts belonging to the target object into local reconstruction of the target object.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image analysis, and in particular to detecting the pose of a target object in an image. Background Art

[0002] In three-dimensional Cartesian space, the position of an object is characterized by three coordinates, and the orientation of the object is characterized by three rotation angles about three axes. The position and orientation constitute the pose of the object, which is also called 6D pose because there are six available degrees of freedom.

[0003] As with many other imaging tasks, machine learning models can be used to detect the pose of a target object in an input image. However, despite achieving satisfactory performance, such machine learning models tend to overfit to specific objects and thus have poor generalization capabilities to unseen objects. Improving generalization capabilities to previously unseen objects requires at least some information about the unseen objects in the best case, and at least partial retraining of the model in the worst case. Summary of the invention

[0004] The present invention provides a method for determining the pose of a target object in an input image. The input image can be acquired using any suitable imaging device, such as a static camera, a video camera, a thermal camera, an ultrasound imaging device, an X-ray imaging device, or a tomographic imaging device.

[0005] The method starts by providing a set of reference images that contain the target object and are annotated with the corresponding pose of the target object. This is the only information about the target object that is needed. In particular, it is not necessary to know the type or size of the target object.

[0006] The reference is input to at least one segmentor. The at least one segmentor is configured to distinguish parts and / or features of the image belonging to different objects. A canonical object model is obtained based on the output of the at least one segmentor obtained for a plurality of reference images. The canonical object model pools information from the reference images in a common coordinate system. That is, the creation of the canonical object model derives information about the hypothesized object in the common coordinate system based on each reference image containing information in a coordinate system different from the common coordinate system.

[0007] In addition, a representation of the target object in the output space of the at least one segmenter is determined based on the output of the at least one segmenter in combination with the annotations of the reference image. That is, the representation contains not only information from the reference image itself, but also ground truth information about the pose of the target object in the reference image.

[0008] The input image is also input to at least one segmenter so that the at least one segmenter generates an output. Based on a comparison of the output on the one hand and a previously obtained representation of the target object on the other hand, it is determined which part of the output of the at least one segmenter obtained for the input image belongs to the target object.

[0009] The parts of the output belonging to the target object are combined into a local reconstruction of the target object.Based on a comparison of the local reconstruction on the one hand and the canonical object model on the other hand, the sought pose of the target object in the input image is determined.

[0010] That is, the determination of which parts of the input image actually belong to the target object is made in the output space of at least one segmenter. In particular, the output space may include a semantic segmentation map that assigns to each pixel of the input image the type of object to which the pixel belongs. But the output space is not limited to this. Alternatively or in combination, the output space may contain richer information, such as a set of features for each image pixel. The local reconstruction is no longer in the output space; instead, it is in the space of the canonical object model. That is, after combining parts of the output of at least one segmenter, a local reconstruction in the space of the object model that is consistent with the combined parts of the output is obtained. The evaluation of the sought pose of the target object is performed in the space of the object model.

[0011] It is found that accurately determining which parts of the input image actually belong to the target object is very important for the accuracy of the final determined pose. In particular, if the input image contains a cluttered scene with many objects, other objects near the target object may have completely different poses. If parts of such other objects are incorrectly determined to belong to the target object, the completely different poses may "poison" the final determined pose of the target object. Removing parts that do not actually belong to the target object from the pose determination greatly improves the accuracy finally obtained. It is found that the proposed analysis in the output space of the segmenter provides a much better separation of the target object from other objects, especially in cluttered scenes where other objects partially occlude the target object.

[0012] In a particularly advantageous embodiment, at least one partitioner comprises:

[0013] First coarse divider It is configured to determine sets of pixels of the image belonging to different objects, and

[0014] Second dense segmenter It is configured to determine a set of features for each pixel of the image.

[0015] In this way, the coarse segmenter The input image can be divided into “patches” belonging to different objects. Then, the dense segmentation can output features that can be used to unambiguously decide which patches belong to the target object. The output features can be used as a “fingerprint” to distinguish the target object from other objects.

[0016] Furthermore, in such an arrangement, the dense partitioner can be trained or fine-tuned in a few-shot manner based on a small number of reference images, while the coarse segmenter Remain frozen in the previously trained state. For example, the coarse segmenter Can be generally trained on large training datasets.

[0017] In a further particularly advantageous embodiment, during the determination of the representation of the target object, for each set of pixels ("patch") belonging to the object according to the coarse segmenter φ, it is determined whether the set of pixels belongs to the target object, based on the annotations of the corresponding reference image. If this is the case, the set of pixels is a "positive patch"; otherwise, the set of pixels is a "negative patch". If the set of pixels is a "positive patch", the representation of the target object represented by the dense segmenter φ is calculated for all pixels in the set. The aggregation (e.g., average) of the output features is used as the positive object-level representation. Similarly, if the set of pixels is a "negative segment", the feature vector generated by the dense segmenter can be calculated for all pixels in the set. The aggregation (e.g., average) of the output features is used as the negative object-level representation.

[0018] The aggregation (e.g., average) of the positive object-level representations is then computed as the representation of the target object. In this way, information from multiple patches belonging to the target object can be aggregated even if these multiple patches are not connected. This may happen, for example, if the target object is partially occluded in the reference image.

[0019] In another particularly advantageous embodiment, the representation dense segmenter is optimized towards the goal of maximizing the pairwise similarity between positive object-level representations The objective may also consist of maximizing the pairwise dissimilarity between positive object-level representations on the one hand and negative object-level representations on the other hand. This makes the optimization objective a “contrastive” objective.

[0020] For example, given M reference images, we can obtain a set of positive target object representations P = {p 1 , ..., p M} and the set N of negative target object representations = {n 1 , ..., n K In this paper, K is random and depends on the number of predicted segments from the reference image. For each positive pair p i and p j , the contrast loss can be calculated as

[0021]

[0022] where τ is a hyperparameter and sim is a similarity measure. The total contrast loss L is calculated over all positive pairs p i and p j Sum and pass Backpropagation is performed to produce updates to its parameters. After the optimization has been completed, the dense segmenter The representation of the target object can be generated by averaging the pixel-wise features over all positive segments.

[0023] In particular, the optimization can be done from the coarse segmenter In this way, the dense partitioner The training of is not done from scratch. Rather, it is simply adapted in a small number of shots based on a small number of reference images.

[0024] In a further particularly advantageous embodiment, the comparison of the output of at least one segmenter comprises a comparison of the output of the at least one segmenter according to the coarse segmenter. The set of each pixel belonging to the object, for all pixels in the set computes the dense segmentation The output features are aggregated as candidate representations. Each such candidate representation is then evaluated for similarity with a previously determined representation of the target object. This can be performed using cosine similarity, for example. The candidate representation for which the similarity is highest is determined to belong to the segmenter of the target object. In this way, given an input image, patches of the input image that are most likely to belong to the target object and that also provide salient information about its pose are selected. In particular, the image quality within the patch also enters into the selection of the most salient patch by measuring the similarity to the representation of the target object in the feature space. For example, if a patch of the input image belongs to the target object but has low quality due to overexposure, underexposure or motion blur, it may be advantageous to select another patch that also belongs to the target object but has better quality.

[0025] In a further particularly advantageous embodiment, in the process of evaluating the sought posture, a region of interest in the input image that is most significant for the evaluation of the sought posture can be determined, at least in part based on the object model and the local reconstruction. This region of interest is then preferred in the evaluation of the sought posture. In this way, the accuracy of the determined posture is improved, because the most reliable information from the input image has the strongest say about the ultimately determined posture. In particular, "preferring" can mean giving a higher weight to information from the region of interest, to the extent that information from input image regions outside the region of interest is not considered at all for the determination of the posture.

[0026] In a further particularly advantageous embodiment, a parameterized rotation R and / or translation T is applied to the object model and / or the local reconstruction. The parameters of the rotation R and / or translation T are optimized towards the goal of improving the match between the object model and the local reconstruction. The optimal rotation R and / or translation T determined in this way can be used to determine an estimate of the pose and / or region of interest. For example, an iterative closest point ICP algorithm can be used to align the object model and the local reconstruction.

[0027] In particular, the region of interest may be determined based at least in part on:

[0028] An estimate of the diameter of the target object derived from the object model

[0029] Optimal translation perpendicular to the image plane as well as

[0030] The focal length f of the camera used to acquire the input image.

[0031] In this way, no prior knowledge about the size of the target object is necessary. and the optimal translation perpendicular to the image plane used as depth information The scale s of the region of interest can be calculated as

[0032]

[0033] Using optimal translation in the image plane and The position [u, v] of the region of interest can be calculated as

[0034] as well as

[0035]

[0036] In a further particularly advantageous embodiment, during the evaluation of the sought pose, a given object detector may be used to compute an image of detection scores in the determined region of interest. The determined pose may then be updated based on the pixel with the highest detection score in the image of detection scores. By limiting the evaluation of the score image to the region of interest, false peaks in the score image are avoided from being misdetected as target objects. In the absence of a region of interest, selecting the highest score from the score image is prone to selecting the wrong target object.

[0037] In a further particularly advantageous embodiment, after the determined pose is updated based on the detection images, the rotation R and / or the translation T are further optimized with the goal of improving the match between the object model and the local reconstruction. In particular, an iterative closest point ICP algorithm can be used again for this purpose. The determined pose can then be updated based on the result of this further optimization. In this way, the tendency of the first optimization of the rotation R and / or the translation T to get stuck in a local optimal 6D pose estimate without a good initialization is avoided.

[0038] In a further particularly advantageous embodiment, an actuation signal is calculated from the determined pose of the target object. The actuation signal is then used to actuate a vehicle, a robot, a driver assistance system, a monitoring system, a quality assurance system and / or a medical imaging system. Since the pose is now determined with better accuracy, the probability is increased that the reaction performed by the actuation technology system in response to the actuation signal is appropriate in the case represented by the input image.

[0039] The method may be computer-implemented in whole or in part and implemented in software. Therefore, the present invention also relates to a computer program having machine-readable instructions, which, when executed by one or more computers and / or computing instances, cause one or more computers and / or computing instances to perform the above method. In this article, control units for vehicles or robots and other embedded systems capable of executing machine-readable instructions are also considered computers. Computing instances include virtual machines, containers, or other execution environments that allow machine-readable instructions to be executed in the cloud.

[0040] The non-transitory storage medium and / or download product may include a computer program. A download product is an electronic product that can be sold online and transmitted over a network for immediate fulfillment. One or more computers and / or computing instances may be equipped with the computer program and / or equipped with the non-transitory storage medium and / or download product.

[0041] Hereinafter, the present invention will be described using the accompanying drawings without any intention of limiting the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 : An exemplary embodiment of a method 100 for determining a pose 2a of a target object 2;

[0043] Figure 2 : A diagram of an exemplary pose estimation pipeline for implementing method 100 including region proposal and refinement;

[0044] Figure 3 : A more detailed illustration of the calculation of the object model 4, the representation 5 of the target object 2 and the local reconstruction 6. DETAILED DESCRIPTION

[0045] Figure 1 is a schematic flow chart of an embodiment of a method 100 for determining a pose 2a of a target object 2 visible in an input image 1 .

[0046] In step 110, a set of reference images 3 containing the target object 2 is provided. These reference images 3 are provided with corresponding poses 2a of the target object 2. * Make a comment.

[0047] In step 120, the reference image 3 is input to at least one segmenter 10, 11, which is configured to distinguish parts and / or features of the image belonging to different objects. In particular, the first segmenter 10 may be a coarse segmenter It is configured to determine sets of pixels of the image belonging to different objects, and the second segmenter 11 may be a dense segmenter It is configured to determine a set of features for each pixel of the image.

[0048] In step 130 , based on the output 10 a , 11 a of at least one segmenter 10 , 11 , a canonical object model 4 is determined which brings together the information from the reference image 3 in a common coordinate system.

[0049] In step 140, the output 10a, 11a of at least one segmenter 10, 11 is combined with the annotation 2a of the reference image 3 * A representation 5 of the target object 2 in an output space of at least one segmenter 10 , 11 is determined.

[0050] According to block 141, for 10 The set of each pixel belonging to the object can be based on the annotation 2a of the corresponding reference image 3 * to determine whether the set of pixels belongs to the target object 2. If this is the case (true value 1), then according to box 142, the dense segmentation can be calculated for all pixels in the set. The aggregation of the output features is taken as the positive object-level representation P. According to block 143 , the aggregation of the positive object-level representations P can then be calculated as the representation 5 of the target object 2 .

[0051] According to block 144, characterize the dense segmenter The parameters of the behavior of 11 can be optimized towards the goal of maximizing the pairwise similarity between the positive object-level representations P. According to block 144a, such optimization can be obtained from the coarse segmentation 10 copies start.

[0052] In step 150, an input image 1 is input to at least one segmenter 10, 11, which produces an output 10b, 11b.

[0053] In step 160 , based on a comparison of said output 10b , 11b of at least one segmenter 10 , 11 with the representation 5 of the target object 2 , it is determined which parts 10c , 11c of said output 10b , 11b of at least one segmenter 10 , 11 belong to the target object 2 .

[0054] According to block 161, when comparing the outputs 10a, 11a of at least one segmenter 10, 11 obtained for the input image 1 and the representation 5 of the target object 2, the outputs 10a, 11a of at least one segmenter 10, 11 obtained for the input image 1 and the representation 5 of the target object 2 are compared. The set of each pixel belonging to an object can be computed for all pixels in the set by the dense segmenter The output features are aggregated as candidate representations7.

[0055] According to box 162 , the similarity between each candidate representation 7 and the representation 5 of the target object 2 is evaluated.

[0056] According to block 163 , the candidate representation 7 for which the similarity is highest can be determined as the segmenter belonging to the target object 2 11 is part of the output.

[0057] In step 170 , parts 10 c , 11 c of the outputs 10 b , 11 b belonging to the target object 2 are combined into a local reconstruction 6 of the target object 2 .

[0058] In step 180 , based on a comparison of the local reconstruction 6 and the canonical object model 4 , a sought pose 2 a of the target object 2 in the input image 1 is estimated.

[0059] According to box 181 , based at least in part on the object model 4 and the local reconstruction 6 , a region of interest 8 in the input image 1 that is most significant for the evaluation of the sought pose 2 a may be determined.

[0060] According to block 181a, the region of interest 8 may be determined based at least in part on:

[0061] An estimate of the diameter of the target object derived from the object model (4)

[0062] Optimal translation perpendicular to the image plane as well as

[0063] The focal length f of the camera used to acquire the input image 1.

[0064] According to box 182, the region of interest 8 may be preferred in the evaluation of the sought pose 2a.

[0065] According to box 183, the parameterized rotation R and / or translation T may be applied to the object model 4 and / or the local reconstruction 6. According to box 184, the parameters of the rotation R and / or translation T may then be optimized towards the goal of improving the match between the object model 4 and the local reconstruction 6.

[0066] According to box 185, if a region of interest 8 has been determined, a given object detector may be used to compute an image 9a of detection scores in the determined region of interest 8. The determined pose 2a may then be updated based on the pixel with the highest detection score in the image of detection scores.

[0067] According to box 187, the rotation R and / or the translation T may then be further optimized towards the goal of improving the match between the object model 4 and the local reconstruction 6. According to box 188, based on the results of this further optimization, the determined pose 2a may be updated.

[0068] In step 190 , an actuation signal 190 a may be calculated based on the determined posture 2 a .

[0069] In step 200 , the vehicle 50 , the robot 51 , the driver assistance system 60 , the monitoring system 70 , the quality assurance system 80 , and / or the medical imaging system 90 are actuated using the actuation signal 190 a .

[0070] Figure 2 is a block diagram of an exemplary pose estimation pipeline for implementing method 100. The pipeline includes an online adaptation module A, a region proposal module B, and a refinement module C.

[0071] The online adaptive module A comprises a first coarse segmenter configured to determine a set of pixels of the image belonging to different objects 10, and a second dense segmenter configured to determine a set of features for each pixel of the image 11.

[0072] When using this combination of segmenters on a reference image 3 with a target object 2, the resulting outputs 10a, 11a can be used to build a canonical object model 4. When using the combination of segmenters on an input image 1, the resulting outputs 10b, 11b can be analyzed with respect to which parts 10c, 11c belong to the target object 2. From these parts 10c, 11c, a local reconstruction 6 of the target object 2 can be constructed. In step 180 of the method 100, based on a comparison of the local reconstruction 6 and the canonical object model 4, a sought pose 2a of the target object 2 in the input image 1 is evaluated.

[0073] To this end, Figure 2 In the region proposal module B in the exemplary pipeline shown in , in steps 183 and 184 of method 100, a parameterized rotation R and / or translation T is applied to the object model 4 and / or to the local reconstruction 6 to achieve optimal alignment between the two. This can be used to predict the region of interest ROI 8 that is most significant for determining the sought pose 2a. Based on this region of interest 8 and the cropped ROI 8 from the reference image 3 * Based on the features of , the Gen6D detector D computes the detection score for image 9. Filtering with ROI 8 produces image 9a. The Gen6D detector D produces a pixel-by-pixel position image that produces a predicted translation T of the target object 2 in the test image 1. At the same time, based on the predicted ROI 8 of the test image 1, the Gen6D selector S predicts the rotation R of the target object 2 in the test image 1. The translation T and the rotation R form an estimate of the pose 2a of the target object 2 in the test image 1.

[0074] In the refinement module C, the estimated pose 2a and the image 9a of detection scores in the ROI 8 are fed to a refiner REF. Based on the pixels with the highest detection scores in this image 9a, the refiner REF produces a refined estimate 2a'. A further refinement 2a" is then obtained by optimizing the rotation R and the translation T using ICP again according to steps 187 and 188 of method 100.

[0075] Figure 3 It is illustrated in more detail how the object model 4 , the representation 5 of the target object 2 and the local representation 6 of the target object 2 are calculated.

[0076] When combined with the pose annotation 2a of reference image 3 * Use coarse segmenter on reference image 3 10 o'clock, coarse divider The output 10a of 10 divides the reference image 3 into negative segments NS that do not belong to the target object 2 and positive segments PS that do belong to the target object 2. Using a dense segmenter The output 11a obtained on the same reference image 3 averages each negative segment NS and each positive segment PS to produce a corresponding negative object-level representation N or positive object-level representation P. Starting with 10 copies, we train the dense segmenter using the contrastive loss function. On the one hand, it is used for the maximum pairwise similarity between positive object-level representations P, and on the other hand, it is used for the maximum pairwise dissimilarity between positive object-level representations P and negative object-level representations N. After training with a few shots on the reference image 3, the dense segmenter 11 produces a positive object-level representation P that can be aggregated (e.g., averaged) into a representation 5 of the target object 2.

[0077] When using the coarse segmenter on input image 1 10, an image 10b of patches belonging to different objects is generated. At the same time, the trained dense segmentation 11 generates dense features for each pixel. These dense features 11b are averaged over each patch from the image 10b to generate a corresponding candidate representation 7. By evaluating the similarity SIM between the candidate representation 7 and the representation 5 of the target object 2, it can be determined which parts 10c, 10c of the segmenter output 10b, 11b belong to the target object 2. From these parts, a local reconstruction 6 of the target object 2 is determined.

Claims

1. A method (100) for determining a pose (2a) of a target object (2) in an input image (1), comprising the following steps: Providing (110) a set of reference images (3) containing the target object (2) and with corresponding poses (2a) of the target object (2) * ) to make comments; inputting (120) a reference image (3) to at least one segmenter (10, 11), said segmenter being configured to distinguish parts and / or features of the image belonging to different objects, The at least one divider (10, 11) comprises: a first rough divider (10), which is configured to determine sets of pixels of the image belonging to different objects; and a second dense segmenter (11) , which is configured to determine a set of features for each pixel of the image; determining (130) a canonical object model (4) based on the output (10a, 11a) of at least one segmenter (10, 11), said canonical object model bringing together information from the reference image (3) in a common coordinate system; Combined with the annotations (2a) of the reference image (3) * ), determining (140) a representation (5) of the target object (2) in the output space of at least one segmenter (10, 11) based on the output (10a, 11a) of at least one segmenter (10, 11), Wherein determining the representation (5) of the target object (2) comprises: For the coarse segmentation (10) The set of each pixel belonging to the object, based on the annotation (2a) of the corresponding reference image (3) * ), determine (141) whether the set of pixels belongs to the target object (2); If the set of pixels belongs to the target object (2), then for all pixels in the set the dense segmentation function is computed (142). The aggregation of the output features is used as the positive object-level representation (P); and Compute (143) the aggregation of positive object-level representations (P) as the representation (5) of the target object (2); inputting (150) an input image (1) to at least one segmenter (10, 11) and determining (160) which parts (10c, 11c) of the resulting output (10b, 11b) of the at least one segmenter (10, 11) belong to the target object (2) based on a comparison of the resulting output (10b, 11b) of the at least one segmenter (10, 11) with a representation (5) of the target object (2); combining (170) the parts (10c, 11c) belonging to the target object (2) into a local reconstruction (6) of the target object (2); and Based on a comparison of the local reconstruction (6) and the canonical object model (4), evaluating (180) a sought pose (2a) of the target object (2) in the input image (1).

2. The method (100) according to claim 1, further comprising: The representation dense segmenter is optimized towards the goal of maximizing the pairwise similarity between positive object-level representations (P) (11) Behavioral parameters.

3. The method (100) of claim 2, wherein the optimization is performed from a coarse segmentation A copy of (10) begins (144a).

4. The method (100) according to any one of claims 1 to 3, wherein the comparison of the output (10a, 11a) of at least one segmenter (10, 11) obtained for the input image (1) with the representation (5) of the target object (2) comprises: · According to the rough segmentation For each pixel belonging to an object, the set (161) of pixels in the set is calculated by the dense segmenter Aggregation of output features as candidate representations (7); evaluating (162) the similarity between each candidate representation (7) and the representation (5) of the target object (2); and Determine (163) the candidate representation (7) for which the similarity is highest as the segmenter belonging to the target object (2) Part of the output of (11).

5. The method (100) according to any one of claims 1 to 4, wherein the evaluation of the sought posture (2a) comprises: determining (181) a region of interest (8) in the input image (1) that is most significant for an evaluation of the sought pose (2a) based at least in part on the object model (4) and the local reconstruction (6); and • Prioritizing (182) the region of interest (8) in the evaluation of the sought pose (2a).

6. The method (100) according to any one of claims 1 to 5, further comprising: Applying (183) the parameterized rotation R and / or translation T to the object model (4) and / or the local reconstruction (6); as well as Optimizing (184) the parameters of the rotation R and / or translation T towards the goal of improving the match between the object model (4) and the local reconstruction (6).

7. The method (100) of claim 5, wherein the region of interest (8) is determined (181a) at least in part based on: An estimate of the diameter of the target object derived from the object model (4) Optimal translation perpendicular to the image plane as well as The focal length f of the camera used to acquire the input image (1).

8. The method (100) according to any one of claims 5 to 7, wherein the evaluation of the sought posture (2a) further comprises: An image 9a of a detection score in the determined region of interest (8) calculated (185) using a given object detector; as well as - Updating (186) the determined pose (2a) based on the pixel with the highest detection score in the image of detection scores.

9. The method according to claims 6 and 8, further comprising: further optimizing (187) the rotation R and / or translation T towards the goal of improving the match between the object model (4) and the local reconstruction (6); and - Updating (188) the determined pose (2a) based on the result of the further optimization.

10. The method (100) according to any one of claims 1 to 9, further comprising: Calculating (190) an actuation signal (190a) based on the determined posture (2a); as well as Actuating (200) a vehicle (50), a robot (51), a driver assistance system (60), a monitoring system (70), a quality assurance system (80), and / or a medical imaging system (90) using an actuation signal (190a).

11. A computer program comprising machine-readable instructions which, when executed by one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100) according to any one of claims 1 to 10.

12. A non-transitory machine-readable data carrier and / or a download product having a computer program as claimed in claim 11.

13. One or more computers and / or computing instances having a computer program as claimed in claim 11 and / or having a machine-readable data carrier and / or a download product as claimed in claim 12.