WO2020239227A1

WO2020239227A1 - A method of rescoring object detections in a digital image

Info

Publication number: WO2020239227A1
Application number: PCT/EP2019/064096
Authority: WO
Inventors: Daniel OLMEDA REINO; Shanshan Zhang; Bernt Schiele
Original assignee: Toyota Motor Europe NV SA; Max-Planck-Institut fur Informatik
Current assignee: Toyota Motor Europe NV SA; Max-Planck-Institut fur Informatik
Priority date: 2019-05-29
Filing date: 2019-05-29
Publication date: 2020-12-03
Anticipated expiration: 2021-11-29

Abstract

A method of rescoring object detections in a digital image (20), wherein an object detection comprises a candidate window indicating the object location and a score representing the confidence of detection, wherein a neural network performs the steps of, for a given object detection (T): - generating a set (24) of other object detections (Pi, PK) which satisfy at least one predetermined criterion with respect to the given object detection (T); - determining a correlation between the given object detection (T) and each element of the set (24); generating a new score for the given object detection (T) based on the correlation.

Description

A METHOD OF RESCORING OBJECT DETECTIONS IN A DIGITAL IMAGE

TECHNICAL FIELD

[0001] The present disclosure relates to object detections in a digital image, and more particularly to a method of rescoring object detections in a digital image, a method of detecting objects in a digital image, a method of training a detector configured to detect objects in a digital image, and related modules and systems. These methods may be used for numerous applications, such as pedestrian detection by vehicles.

TECHNOLOGICAL BACKGROUND

[0002] Pedestrian detection is an important problem in computer vision, serving as a basic building block in many intelligent systems such as autonomous driving and video surveillance. During the past few years, the performance has been improved largely due to the development of convolutional neural networks (shortened as convnets or CNNs). Although the progress is significant across different datasets, performance is still low for crowded scenes. Importantly, partial occlusion remains a major challenge for pedestrian detection, occurring frequently in practice. Recently, some new datasets consisting of dense people crowds have been proposed to encourage more work investigating occlusion handling, such as CityPersons (Zhang, Shanshan, Rodrigo Benenson, and Bernt Schiele, "CityPersons: A Diverse Dataset for Pedestrian Detection. " 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)).

[0003] There are several approaches focusing on occlusion handling. The key problem is how to model various occlusion patterns in an effective way. Most approaches pre-define occlusion patterns with a combination of a rectangular visible box and a full body box, and then learn corresponding representations for each pattern. However, one single rectangular box is not able to cover the entire visible area when the occlusion is irregular, especially introduced by a nearby human body. Using a plurality of rectangular boxes to cover the visible parts of a human body is expensive to annotate, and it is also unclear how effective such annotations could be employed.

[0004] More recently, it has been proposed to use part detections as fine-grained guidance to allow the representations to focus more on visible parts in a dynamic way. However, when the occlusion is caused by another person, the part detections from the occluder tend to mix up with those from the occluded person, thus resulting in confusion.

[0005] Thus, from the analysis of crowded scenes, it is unclear how either the visible box or part detections can be used to identify irregular and challenging person-person occlusions. Therefore, there is a need for a new method in order to better detect objects, in particular handle the irregular person-person occlusion patterns.

SUMMARY

[0006] In this respect, the present disclosure relates to a method of rescoring object detections in a digital image, wherein an object detection comprises a candidate window indicating the object location and a score representing the confidence of detection, wherein a neural network performs the steps of, for a given object detection:

- generating a set of other object detections which satisfy at least one predetermined criterion with respect to the given object detection;

- determining a correlation between the given object detection and each element of the set;

- generating a new score for the given object detection based on the correlation. [0007] Owing to the use of a neural network, this rescoring method is generally a computer implemented method. The neural network may be of any type, e.g. a CNN.

[0008] The digital image, the candidate window and the (initial) score may be obtained by a previously applied method. The candidate window may indicate the object location in the digital image. The candidate window may be a box, such as a so-called bounding box bounding the shape of the object in the image. The score may indicate the confidence with which the candidate window corresponds to a sought object.

[0009] The other object detections are preferably selected in the same digital image. Each one of the other object detections may also comprise a candidate window and a score.

[0010] In generating a set of other object detections as recited above, the rescoring method generates pairs, each pair comprising the given object detection and an element of the set.

[0011] In generating a set of other object detections as recited above, the method sorts the object detections and select those who satisfy at least one predetermined criterion with respect to the given object detection. The criterion may be chosen so that the correlations between the given object detection and the elements of the set, then determined, provide relevant information about the given object detection. For instance, if the set is generated to include object detections looking like a sought object, the correlations will indicate whether the given object detection looks like the sought object. A new score for the given object detection is generated accordingly.

[0012] Thus, by exploiting relations and correlations between objects, the above method provides an effective way to accurately rescore object detections in a digital image. As a consequence, the confidence of detection is improved. [0013] Optionally, the given object detection is a pedestrian detection. That is, before carrying out the rescoring method, the given object detection is believed to be a pedestrian detection, although the rescoring method may rescore the given object detection so that it turns out that the detected object is not a pedestrian. A pedestrian detection has a candidate window supposed to indicate the location of a pedestrian, and a score that is high when it is highly probable that the object is a pedestrian, and low when it is highly probable that the object is not a pedestrian. Selecting a score threshold for an object detection to qualify as a pedestrian detection is within the skilled person's abilities.

[0014] Optionally, the other object detections are pedestrian detections. The same remarks as above apply likewise.

[0015] Optionally, the at least one predetermined criterion comprises at least one of an intersection over union ratio of the candidate window of the given object detection and the candidate window of the element being greater than a predetermined value, and the score of the element being greater than a predetermined value.

[0016] The intersection over union (IoU) ratio, as known in the art per se, is an example for measuring the degree of overlapping of the candidate window of the given object detection and the candidate window of the element. In these embodiments, the other object detections are retained in the set if they sufficiently overlap with the given object detection. Thus, the given object detection and the element of the set are likely to constitute a pair of occluder object/occluded object. Accordingly, detection of occluded objects is improved. Selecting an IoU threshold for an object detection to be retained in the set is within the skilled person's abilities.

[0017] Alternatively or complementarily, by selecting other object detections which have a score greater than a predetermined value, the set is built to include object detections which are likely to be the sought objects (e.g., which are likely to be pedestrians). Thus, when determining a correlation between the given object detection and an element of the set, the correlation is likely to give better results because the set elements have been appropriately selected. Accordingly, detection of occluded objects is improved.

[0018] Optionally, the correlation is determined at least based on an appearance similitude between the given object detection and the element. That is, the correlation is high when the given object detection "looks like" the element of the set, and otherwise low, so that rescoring can be carried out accordingly. It should be noted that the appearance may be determined based on the visible content of the object detection, or on features determined from the object detection.

[0019] For instance, optionally, the appearance similitude is determined based on a latent representation of the given object detection and a latent representation of the element. The latent representation may be calculated by the rescoring method for this purpose, or obtained from elsewhere, e.g. from a previous application of another method, such as the previous application of a method generating the object detection.

[0020] A latent representation may be a multi-dimensional feature vector. For example, latent representation may be determined based on the candidate window, and in particular image content in the candidate window. An example would be to take the image content in the window and process it by at least one (or e.g. three) layer of a neural network, in order to obtain the latent representation/feature vector. In a specific example the at least one layer consists of at least one of a convolution, a pooling, and/or a FC (fully connected) layer.

[0021] Optionally, the correlation is determined at least based on the score of the element. For instance, the correlation may be high if the score of the element is high, and otherwise low, so that rescoring can be carried out accordingly. [0022] Optionally, generating a new score comprises generating a relation feature pool based on the correlation between the given object detection and each element of the set, and reweighting the score based on the relation feature pool. Thus, the initial score is taken into account and corrected on the basis of the relation feature pool.

[0023] The present disclosure is further directed to a method of detecting objects in a digital image, comprising the steps of:

- obtaining at least one object detection from the digital image, wherein the object detection comprises a candidate window indicating the object location and a score representing the confidence of detection,

- generating a new score of the object detection by the previously described method of rescoring, and

- returning the candidate window and the new score.

[0024] As above, this detecting method is generally a computer implemented method.

[0025] The obtaining step may be performed in several ways. For instance, the method may include an actual detection step, taking the digital image as an input and returning the at least one object detection. This detection step may use a separate (and potentially carried out previously) or integrated detection method.

[0026] Thanks to the method of rescoring, the present method of detecting objects provides object detections with improved scores.

[0027] The present disclosure is further directed to a method of training a detector configured to detect objects in a digital image, the method comprising the steps of:

- inputting a digital image to the detector,

- obtaining at least one object detection returned by the detector, wherein the object detection comprises a candidate window indicating the object location and a score representing the confidence of detection,

- updating the detector based on the new score.

[0028] The detector may comprise a detector model, i.e. a mathematical representation configured to detect objects in a digital image. For instance, the detector may include a neural network, e.g. a CNN. The method of training may be a computer implemented method. Detectors are known per se in the art, e.g. from Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, Faster R-CNN: towards real-time object detection with region proposal networks, in Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1 (NIPS’15).

[0029] The updating may include modifying a score generating function of the detector, e.g. weights thereof, so that the score generating function directly includes the possible corrections brought by the method of rescoring. Afterwards, the detector may be used on its own, without the method of rescoring, and yet may provide improved results as compared with a similar detector which has not undergone this training.

[0030] The present disclosure is further directed to a module for rescoring object detections in a digital image, wherein an object detection comprises a candidate window indicating the object location and a score representing the confidence of detection, wherein the module comprises a neural network configured, for a given object detection, to:

- generate a set of other object detections which satisfy at least one predetermined criterion with respect to the given object detection;

- determine a correlation between the given object detection and each element of the set; - generate a new score for the given object detection based on the correlation.

[0031] The module may implement the above-described method for rescoring. The module may have the hardware structure of a computer.

[0032] The present disclosure is further directed to a detection system for detecting objects in a digital image, comprising a detector configured to obtain at least one object detection from the digital image, the object detection comprising a candidate window indicating the object location and a score representing the confidence of detection, the detection system further comprising the module for rescoring as previously described, the detection system being configured to return the candidate window obtained by the detector and the new score generated by the module for rescoring.

[0033] The detection system may implement the above-described method of detecting. The detection system may have the hardware structure of a computer.

[0034] Optionally, the detection system is further configured to update the detector based on the new score.

[0035] Optionally, the detector comprises a neural network.

[0036] The present disclosure is further directed to a vehicle comprising the module as previously described or the detection system as previously described.

[0037] The present disclosure is further directed to a computer program comprising instructions for executing the steps of any one of the previously described methods when the program is executed by a computer.

[0038] This program can use any programming language and take the form of source code, object code or a code intermediate between source code and object code, such as a partially compiled form, or any other desirable form. [0039] According to another aspect of the embodiments of the present disclosure, there is provided a recording medium readable by a computer and having recorded thereon a computer program including instructions for executing the steps of a method as defined above.

[0040] The recording medium can be any entity or device capable of storing the program. For example, the medium can include storage means such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or magnetic storage means, for example a diskette (floppy disk) or a hard disk.

[0041] Alternatively, the recording medium can be an integrated circuit in which the program is incorporated, the circuit being adapted to execute the method in question or to be used in its execution.

BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The invention and advantages thereof will be better understood upon reading the detailed description which follows, of embodiments given as non-limiting examples. This description refers to the appended drawings, wherein:

- Fig. 1 is a block diagram representing a detection system according to an embodiment;

- Fig. 2 is a diagram showing a vehicle including a detection system according to an embodiment;

- Fig. 3 shows an example of a candidate window in a digital image;

- Fig. 4 shows candidate windows of a set of other object detections in the digital image of Fig. 3;

- Fig. 5 illustrates a correlation between two object detections in the digital image.

DETAILED DESCRIPTION OF EMBODIMENTS

[0043] Figure 1 shows a block diagram of a detection system 10 comprising an object detector 22 (detector 22) and a rescoring module 12 (i.e. a module for rescoring object detections) according to embodiments of the present disclosure. The system may have various further functions, e.g. may be a robotic system or a camera system. It may further be integrated in a vehicle 50, as illustrated in Fig. 2.

[0044] The detection system 10 may comprise an electronic circuit, a processor (shared, dedicated, or group), a combinational logic circuit, a memory that executes one or more software programs, and/or other suitable components that provide the described functionality. In other words, the detection system 10 may be a computer device. The detection system 10 may be connected to a memory, which may store data, e.g. a computer program which when executed, carries out the method according to the present disclosure. In particular, the detection system or the memory may store software which comprises the neural network according to the present disclosure.

[0045] The detection system 10, in particular the detector 22, has an input for receiving a digital image 20 or a stream of digital images. In particular, the detection system 10, in particular the detector 22, may be connected to an optical sensor, in particular a digital camera. The digital camera is configured such that it can record a scene, and in particular output digital data to the detection system 10, in particular the detector 22.

[0046] The detector 22 may be implemented as software running on the detection system 10 or as a hardware element of the detection system 10. The detector 22 carries out a computer vision algorithm for detecting the presence and location of objects in a sensed scene. For example, vehicles, persons (pedestrians), and other objects may be detected. The detector 22 outputs at least one candidate window indicating an object location and generates for each candidate window a score representing the confidence of detection.

[0047] Furthermore, also the rescoring module 12 may be implemented as software running on the detection system 10 or as a hardware element of the detection system 10. In particular, the detection system 10 may comprise a neural network which includes both the detector 22 and the rescoring module 12. Alternatively the rescoring module 12 may be realized as an independent neural network (in particular beside a neural network comprising the detector 22).

[0048] The rescoring module 12 receives as an input from the detector the detection results. In particular, it receives information regarding one or more object detections T. Each object detection T comprises a candidate window indicating the object location and a score representing the confidence of detection. The rescoring module 12 rescores these objects detections T such that confidence of detection is improved. In detail, for a given object detection T, the rescoring module 12 generates a set 24 of other object detections P₁, P₂, ..., P_K which satisfy at least one predetermined criterion with respect to the given object detection T. A correlation is then determined between the given object detection T and each element P, of the set 24, and the rescoring module 12 generates a new score for the given object detection T based on the correlation. The detection system 10, and particularly the rescoring module 12, is able to output an updated object detection T, comprising the candidate window of the given object detection and the new score, i.e. the re-evaluated score.

[0049] By this, confidence of the detection is improved. As will be described, this method can be advantageously used for detection of occluded objects, e.g. occluded persons.

[0050] In addition, the new score may be used to update the detector 22, as shown in Fig. 1. Training of the detector 22 based on a new score for the given object detection is known perse in the art.

[0051] In the following, the operation of the neural network according to an embodiment the present disclosure is explained in more detail and with reference to figures 3 to 5, focusing on an application to pedestrian detection.

[0052] A major challenge of handling crowds comes from the irregular occlusion patterns, which lead to rather diverse appearances of human bodies. In this disclosure, it is proposed to employ relation features in convnets allowing the network to produce higher confidence score for occluded ones.

[0053] Since the Adapted FasterRCNN detector from Zhang, Shanshan, Rodrigo Benenson, and Bernt Schiele, "CityPersons: A Diverse Dataset for Pedestrian Detection. " 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017), achieves state-of-the-art performance, it is used herein as the base detector, namely the detector 22, while adding a relation module (rescoring module 12) as a refinement procedure for rescoring. As described above, the flowchart of the approach is shown in Fig. 1. For each proposal detection box (candidate window of a given object detection), relation pairs (i.e. pairs comprising the given object detection T and an element P, of the set 24) are first generated based on a predetermined criterion, e.g. the relative position or original scores; then for each relation pair, features may be extracted and concatenated to generate a large relation feature pool F; afterwards, these relation features may be passed into a linear transformer 28 to produce a relation score; finally, the new score of the updated object detection T' is calculated, e.g. by multiplying the original and relation scores.

[0054] As shown in Fig. 1, relation pairs are generated. For each occluded person T (target proposal), ideally it is desired to find some high scoring proposals that are highly correlated with it, so as to guide the classifier to produce a high rescoring weight. To this end, as an example, two different strategies are proposed, in alternative to each other or in addition to each other, to sample relation proposals within a local neighborhood of a proposal: (1) IoU (Intersection over Union) priority: those proposals which have high overlap ratios with the target proposal are sampled; in other words, an element is retained in the set 24 if an IoU ratio of the candidate window of the given object detection and the candidate window of the element is greater than a predetermined value. (2) Score priority: those proposals which have high scores are sampled; in other words, an element is retained in the set 24 if the score of the element is greater than a predetermined value.

[0055] An example is shown in Figs. 3 and 4. Here, the target proposal T is denoted by a box in the digital image 20 of Fig. 3; the proposals sampled based on IoU priority and score priority are shown respectively on the top and bottom of Fig. 4. From these illustrations, it can be seen that IoU priority has a larger chance to sample person proposals, while score priority may sample some high scoring false positives from the background, for example, the street light (see proposal P_K in Fig. 4). The inventors have found that in general, IoU priority is better than score priority sampling: IoU priority has a higher chance to sample nearby persons in crowds, while score priority may result in sampling on some high scoring background objects (even though the score priority strategy is also restricted to a local neighborhood around the proposal).

[0056] Thus, for each target proposal T (given object detection T), K proposals P₁ to P_K are sampled using at least one of IoU and score priority and relation pairs R are then generated as follows:

[0057] Generating these pairs is equivalent to generating the set 24 of other object detections, the set 24 being

The inventors have found that the number of relation pairs (i.e. number K) has a relative small effect on the performance. The reason is that there usually exists a large number of person proposals in crowded scenes.

[0058] For each relation pair two kinds of features are

computed: (1) appearance correlation (or appearance similitude) features; (2) rank features. Computing the features for a pair corresponds to the above-mentioned determining a correlation.

[0059] Regarding the appearance features, the correlation is determined at least based on an appearance similitude between the given object detection (target proposal) T and the element of the set (proposal) Pi. The appearance features for each proposal are picked from the last convolutional layer of the base detector 22. More generally, the appearance similitude may be determined based on a latent representation of the given object detection and a latent representation of the element. In the present example, the latent representation is that provided by the last convolutional layer of the detector 22. However, the latent representation may be obtained differently, e.g. from another layer and/or through another transformation of the digital image.

[0060] According to the inventors' analysis, the subregions of the candidate windows show high correlation between the occluded person and the person in front, and thus the correlation based on an appearance similitude is determined by computing the Hadamard product of the two feature arrays as the appearance correlation features:

where f_A(T) and f_A(P_i) are pooled appearance features for T and P_i, respectively; and ° indicates element-wise product.

[0061] A diagrammatic visualization example of appearance correlation features are shown in Fig. 5, based on the target proposal T of Fig. 3 and the first proposal P₁ of Fig. 4. Fig. 5 shows the appearance correlation features across multiple representative channels. As can be clearly seen, the appearance correlation features show rather high values (see areas 32 in Fig. 5) at visible parts.

[0062] To reduce the dimension of features, a 3 x 3 max pooling may be performed on top of the correlation features for each channel.

[0063] In addition or as an alternative to appearance correlation features, the rank of P_i may also be considered. Thus, the correlation is determined at least based on the score of the element P_i. Proposal T has a high probability of being a person (or, more generally, a sought object) when it has high correlation with high scoring proposals (i.e. proposals likely to be a sought object). In other words, the correlation is determined at least based on the score of the element P_i. Here, the rank is computed by taking into account all proposals in one image, rather than selected relation proposals. The one dimensional rank value is embedded into a high dimensional representation using the method of A. Vaswani, N. Shazeer, N. Pamar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser and I. Polosukhin, Atention is all you need, in NIPS 2017, which computes cosine and sine functions of different frequencies: and

where d is the rank feature dimension after embedding; r(P_i) is the rank value for proposal i, which is an operation of linear algebra known per se in the art; t is the dimension.

[0064] The relation feature pool F for proposal T is generated by concatenating the appearance correlation features and rank features, as applicable, for all pairs:

The usage of rank features together with appearance correlation features ensures that those target proposals containing a person are assigned higher weights than background proposals. [0065] The relation features are passed through a linear transformer 28 to regress the rescoring weight of the target proposal T, denoted as wgt(T):

The linear transformer 28, represented by the function G, may be a scalar product with a vector configured to give high scores to a desired object, e.g. pedestrians.

[0066] Then the refinement is implemented by re-weighting the original score:

where score_ori(T ) is the original or initial score of proposal T from the base detector 22.

[0067] The refinement network, namely the rescoring module 12, is trained by optimizing the cross-entropy loss for classification. After training, the classifier is expected to output a higher score for occluded persons, while suppressing proposals from the background.

[0068] An implementation example is now described. An average- log miss rate (MR) is used in all experiments, which is computed in the false positive per image (FPPI) range of [10^-2, 10°]. In a context of detecting occluded pedestrians, the results will be shown across different occlusion levels:

(1) Reasonable (R): visibility Î [0.65, inf];

(2) Heavy occlusion (HO): visibility Î [0.20,0.65];

(3) Reasonable + Heavy occlusion (R+HO): visibility Î [0.20, inf].

[0069] The performance on the R+HO subset is used to measure the overall performance as it includes a wide range of occlusions. Note that only pedestrians with height greater than or equal to 50 cm are considered. Piotr Dollar's matlab evaluation code, available at

http://www.vision.caltech.edU/Image_Datasets/CaltechPedestrians/code/c ode3.2.1.zip, is used for evaluation. [0070] The CityPersons dataset, described in the above-mentioned paper and used in this example, shows high diversity of images. More importantly, it includes a large number of crowded scenes as the data was recorded in the center of many big cities, e.g. Cologne and Frankfurt. Here, the original training and validation split, which are composed of 2,975 and 500 images respectively, are used. The 500 images of the CityPersons dataset have 1579 pedestrians in the R subset, 733 pedestrian in the HO subset, which leads to a total of 2312 pedestrian in the R+HO subset, i.e. a density of 4.62 pedestrians per image.

[0071] Two different base detectors are used for the examples: FasterRCNN (Ren et al., cited above) and RFCN (K. H. J. S. Jifeng Dai, Yi Li, R-FCN: object detection via region-based fully convolutional networks, in NIPS, 2016). Modifications are made to both detectors following Zhang et al., cited above. The base detector parameters are fixed while training the relation module. All proposals (-2000) from the base detector without non-maximum suppression (NMS) are sent to the rescoring module 12 for refinement. A standard NMS (IoU threshold = 0.5) is then applied on the refined proposals.

[0072] On the CityPersons dataset, the original scale lx of the input images is used, as it is more than 2x faster for both training and testing, resulting in only a small performance drop of -lpp. The parameters are initialized using a Gaussian filler. The training is performed with an initial learning rate of 10^-3 for 20,000 iterations and for another 40,000 iterations with a decreased learning rate of 10^-4.

[0073] In these examples, only the IoU priority is used for sampling relation pairs (i.e. generating the set), and the parameters are set to K = 5, d = 64. The dimension of final features by concatenating appearance and relation features is 4672.

[0074] The results obtained on the CityPersons dataset are as follows. The efficacy of the proposed method is first validated by applying it on top of two baseline detectors: FasterRCNN and RFCN. Both baselines are modified following the practice from above-cited Zhang et al, to better adapt to the pedestrian detection task. The results on the CityPersons validation set are shown in the Table below, wherein numbers are the miss rates. As can be seen, the relation module (rescoring module 12) improves over the baselines on the heavy occlusion subset by a large margin, more than 8pp for FasterRCNN and more than 9pp for RFCN. The method brings a large gain for person-person (P-P) occlusions; however, even for persons occluded by other objects (P-O), the method is helpful as more context is used to enrich the features. These improvements demonstrate that the relation module is able to recover those occluded persons through the refinement procedure.

[0075] The FasterRCNN+relation detector, as an embodiment of the detection system 10, achieves better results, or at least comparable results, than state-of-the-art methods, including methods using extra cues and annotations, e.g. visible boxes or part detections.

[0076] The tested example embodiment also provides very good results on other datasets, including datasets having a density of pedestrian per image ranging from less than 1 to more than 20.

[0077] Although the present disclosure refers to specific exemplary embodiments, modifications may be provided to these examples without the departing from the general scope of the invention as defined by the claims. In particular, individual characteristics of the different illustrated/mentioned embodiments may be combined in additional embodiments. Therefore, the description and the drawings should be considered in an illustrative rather than in a restrictive sense.

Claims

1. A method of rescoring object detections in a digital image (20), wherein an object detection comprises a candidate window indicating the object location and a score representing the confidence of detection, wherein a neural network performs the steps of, for a given object detection (T):

- generating a set (24) of other object detections (P₁, ..., P_K) which satisfy at least one predetermined criterion with respect to the given object detection (T);

- determining a correlation between the given object detection (T) and each element of the set (24);

- generating a new score for the given object detection (T) based on the correlation.

2. The method of claim 1, wherein the given object detection is a pedestrian detection.

3. The method of claim 1 or 2, wherein the other object detections are pedestrian detections.

4. The method of any one of claims 1 to 3, wherein the at least one predetermined criterion comprises at least one of an intersection over union ratio of the candidate window of the given object detection and the candidate window of the element being greater than a predetermined value, and the score of the element being greater than a predetermined value.

5. The method of any one of claims 1 to 4, wherein the correlation is determined at least based on an appearance similitude between the given object detection and the element.

6. The method of claim 5, wherein the appearance similitude is determined based on a latent representation of the given object detection and a latent representation of the element.

7. The method of any one of claims 1 to 6, wherein the correlation is determined at least based on the score of the element.

8. The method of any one of claims 1 to 7, wherein generating a new score comprises generating a relation feature pool (26) based on the correlation between the given object detection (T) and each element of the set (24), and reweighting the score based on the relation feature pool.

9. A method of detecting objects in a digital image (20), comprising the steps of:

- obtaining at least one object detection (T) from the digital image (20), wherein the object detection (T) comprises a candidate window indicating the object location and a score representing the confidence of detection,

- generating a new score of the object detection by the method of any one of claims 1 to 8, and

- returning the candidate window and the new score (T).

10. A method of training a detector (22) configured to detect objects in a digital image, the method comprising the steps of:

- inputting a digital image (20) to the detector (22), - obtaining at least one object detection (T) returned by the detector (20), wherein the object detection (T) comprises a candidate window indicating the object location and a score representing the confidence of detection,

- updating the detector based on the new score.

11. A module (12) for rescoring object detections in a digital image (20), wherein an object detection comprises a candidate window indicating the object location and a score representing the confidence of detection, wherein the module (12) comprises a neural network configured, for a given object detection (T), to:

- generate a set (24) of other object detections ( P₁, ..., P_K) which satisfy at least one predetermined criterion with respect to the given object detection (T);

- determine a correlation between the given object detection (T) and each element of the set (24);

- generate a new score for the given object detection (T) based on the correlation.

12. A detection system (10) for detecting objects in a digital image (20), comprising a detector (22) configured to obtain at least one object detection (T) from the digital image, the object detection comprising a candidate window indicating the object location and a score representing the confidence of detection, the detection system further comprising the module for rescoring (12) of claim 11, the detection system (10) being configured to return the candidate window obtained by the detector and the new score generated by the module for rescoring (T').

13. The detection system of claim 12, being further configured to update the detector (22) based on the new score.

14. A vehicle (50) comprising the module for rescoring (12) as claimed in claim 11 or the detection system (10) as claimed in claim 12 or 13.

15. A computer program comprising instructions for executing the steps of the method of claim 1, claim 9 or claim 10 when the program is executed by a computer.