Information processing apparatus, information processing method, and program

Abstract

To suppress the occurrence of erroneous detection of another subject similar to a subject to be tracked.SOLUTION: Acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject, and acquiring a first feature amount extracted from the first subject of the first image and a second feature amount extracted from the second subject of the second image; Detecting a first likelihood of the first subject in the third image based on the third image and the first feature amount, and detecting a second likelihood of the second subject in the third image based on the third image and the second feature amount; Then, using a loss function for acquiring a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood, the extraction model used for feature extraction and the detection model used for likelihood detection are updated to reduce the acquired loss.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing technique for tracking a subject in an image. [Background technology]

[0002] There are techniques that use deep neural networks to track subjects in videos. One method that uses deep neural networks is to use a correlation filter, as disclosed in Patent Document 1 and Non-Patent Document 1. In this method, correlation features are acquired from the image region of the subject to be tracked when tracking begins, and in processing subsequent images in the video, the correlation features are used as a convolution kernel to detect the subject to be tracked. This method makes it possible to accurately detect the subject to be tracked by estimating areas in the image that have a high correlation with the correlation features of the image region of the subject. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2021-526269 [Non-patent literature]

[0004] [Non-Patent Document 1] Luca Bertinetto, Jack Valmadre, Joao F. Henriques, Andrea Vedaldi, Philip HS Torr:Fully-Convolutional Siamese Networks for Object Tracking(ECCV2016) Summary of the Invention [Problem to be solved by the invention]

[0005] In the case of the method of tracking a subject using the aforementioned correlation features, in cases where multiple subjects with similar appearances co-occur or intersect, such as in sports or horse racing, a subject that looks similar to the subject being tracked may be mistakenly detected as the subject being tracked.

[0006] Therefore, an object of the present invention is to make it possible to suppress the occurrence of erroneous detection of other subjects that are similar to the subject to be tracked. [Means for solving the problem]

[0007] The information processing device of the present invention is characterized by comprising: an acquisition means for acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction means for acquiring a first feature extracted from the first subject in the first image and a second feature extracted from the second subject in the second image; a detection means for detecting a first likelihood in the third image for the first subject based on the third image and the first feature, and for detecting a second likelihood in the third image for the second subject based on the third image and the second feature; a loss calculation means using a loss function for obtaining a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; and an update means for updating an extraction model used by the extraction means to extract the feature and a detection model used by the detection means to detect the likelihood, so as to reduce the loss obtained by the loss calculation means. [Effects of the Invention]

[0008] According to the present invention, it is possible to suppress the occurrence of erroneous detection of other subjects similar to the subject to be tracked. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 2 is a diagram illustrating an example of a functional configuration of a learning device. [Figure 2] 4 is a flowchart of a learning process according to the first embodiment. [Figure 3] FIG. 2 is a diagram showing the relationship between pieces of information in the learning device. [Figure 4] 10 is a flowchart of a learning process according to a second embodiment. [Figure 5] FIG. 10 is a diagram showing extraction positions of subject feature vectors. [Figure 6] FIG. 2 is a diagram illustrating an example of a functional configuration of an inference device. [Figure 7] 10 is a flowchart of an inference process according to the third embodiment. [Figure 8] FIG. 2 illustrates an example of a hardware configuration. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The following embodiments do not limit the present invention, and not all of the combinations of features described in the present embodiments are necessarily essential to the solution of the present invention. The configurations of the embodiments may be appropriately modified or changed depending on the specifications of the device to which the present invention is applied and various conditions (such as usage conditions and usage environment). Furthermore, the configurations of each embodiment may be appropriately combined. In the following embodiments, redundant descriptions of the same or similar configurations and processing steps will be omitted. In the following embodiments, positions on an image and positions on a feature plane (a two-dimensional extracted feature plane) when feature amounts extracted by a feature amount extraction model of a deep neural network are stored are represented by (x, y) coordinates with the upper left pixel as the origin (0, 0). In addition, in this embodiment, the subject to be tracked is referred to as a tracked subject.

[0011] <First embodiment: learning device> First, as a first embodiment, we will explain an information processing device (hereinafter referred to as a learning device) that learns a deep neural network model for tracking a tracking subject within a series of multiple images (still images per frame) that are sequential in time, such as a video. FIG. 1 is a diagram showing an example of the functional configuration of a learning device 100 according to this embodiment.

[0012] The acquisition unit 110 of the learning device 100 shown in FIG. 1 acquires a learning data set to be used in the learning process. In this embodiment, the training dataset is composed of a training image set, in which a plurality of time-series images (still images) are prepared as training images (hereinafter referred to as training images), and correct answer data corresponding to each training image in the training image set. Each training image in the training image set contains at least one subject (object), including a tracking subject. In this embodiment, the time-series training images are assumed to be, for example, images of each frame constituting a video captured by a video capture device, but may also be multiple images generated by an image generation device. The correct answer data is in one-to-one correspondence with each training image in the training image set and is data including at least position information of each subject in the training image. The position information of each subject is assumed to be represented by the central coordinates of the area of ​​each subject in the training image. The acquisition unit 110 may acquire one training dataset, or two or more different training datasets.

[0013] The acquisition unit 110 acquires, from the training dataset, at least three training images, namely, a first training image, a second training image, and a third training image, as well as first correct answer data, second correct answer data, and third correct answer data that correspond one-to-one to each of the training images. In this embodiment, the training image set from which the second training images are acquired may be the same training image set from which the first training images were acquired, or may be a different training image set from the training image set from which the first training images were acquired. Furthermore, the second training image may be the same image as the first training image. The training image set from which the third training images are acquired is the same training image set from which the first training images were acquired.

[0014] Acquisition unit 110 also cuts out an image area of ​​the tracking subject from the first learning image. Hereinafter, the image area of ​​the tracking subject cut out from the first learning image is referred to as the tracking subject image. Acquisition unit 110 also cuts out an image area of ​​a subject other than the tracking subject from the second learning image. Hereinafter, the subject other than the tracking subject is referred to as the non-tracking subject, and the image area of ​​the non-tracking subject cut out from the second learning image is referred to as the non-tracking subject image. Acquisition unit 110 also cuts out an image area of ​​the tracking subject from the third learning image. Hereinafter, the image area of ​​the tracking subject cut out from the third learning image is referred to as the target tracking subject image.

[0015] Furthermore, the acquisition unit 110 creates a positive correct likelihood map and a negative correct likelihood map based on the third correct answer data. In this embodiment, the positive correct answer likelihood map based on the third correct answer data is a map in which the position information of the tracking subject indicated by the third correct answer data is replaced with a likelihood distribution, and the negative correct answer likelihood map is a map in which all pixel values ​​are the minimum value of 0.

[0016] Extraction unit 120 has a feature extraction model of a deep neural network, and extracts image feature amounts from each of the tracked subject image and the non-tracked subject image. Hereinafter, image feature amounts extracted from the tracked subject image will be referred to as tracked subject feature amounts, and image feature amounts extracted from the non-tracked subject image will be referred to as non-tracked subject feature amounts.

[0017] The object detection unit 130 uses a tracking object detection model of a deep neural network to acquire a first likelihood based on the target tracking object image and tracking object feature amounts, and acquires a second likelihood based on the target tracking object image and non-tracking object feature amounts. In this embodiment, the object detection unit 130 acquires a map representing the first likelihood and a map representing the second likelihood. In this embodiment, the map representing the first likelihood is a map in which each pixel indicates the likelihood of a tracking object, and is the target of learning (learning using a positive loss) to ensure that pixels corresponding to tracking objects output high values. Hereinafter, the map representing the first likelihood will be referred to as a positive object likelihood map. The map representing the second likelihood is a map in which each pixel indicates the likelihood of a tracking object, similar to the map representing the first likelihood, but is the target of learning (learning using a negative loss) to ensure that the minimum value (0) is output for all pixels. Hereinafter, the map representing the second likelihood will be referred to as a negative object likelihood map. The positive object likelihood map and the negative object likelihood map are maps of arbitrary width and height. In addition to the positive object likelihood map, object detection section 130 may output a size map indicating the size of the tracked object, etc.

[0018] The loss calculation unit 140 performs a loss calculation process to calculate a first loss based on the positive ground truth likelihood map and the positive object likelihood map, and to calculate a second loss based on the negative ground truth likelihood map and the negative object likelihood map. Details will be described later, but in this embodiment, the loss calculation unit 140 performs a process to acquire a loss such that the loss is smaller when the first loss is larger than the second loss, and the loss is larger when the second loss is larger than the first loss. Hereinafter, in this embodiment, the first loss is referred to as a positive loss, and the second loss is referred to as a negative loss. That is, the loss calculation unit 140 uses a loss function that acquires a loss such that the loss is smaller when the positive loss is larger than the negative loss, and the loss is larger when the negative loss is larger than the positive loss.

[0019] The weight update unit 150 updates the parameters of the feature extraction model of the extraction unit 120 and the parameters of the tracking object detection model of the object detection unit 130 so as to reduce the sum of the losses calculated by the loss calculation unit 140. For example, backpropagation can be used to update the weight parameters.

[0020] FIG. 2 is a flowchart showing the information processing performed by the learning device 100 of this embodiment, that is, the processing flow for learning a feature extraction model and a tracking object detection model. The learning processing according to this embodiment is performed by repeating the processing shown in the flowchart of FIG. 2 a desired number of times. FIG. 3 is a diagram showing the relationship between each piece of information in the learning device 100 of this embodiment. The learning processing performed by each functional unit of the learning device 100 of this embodiment will be described below with reference to FIGS. 1 to 3.

[0021] In the learning device 100 of this embodiment, the acquisition unit 110 acquires a learning dataset 300 to be used in the learning process. As described above, the learning dataset 300 is composed of a training image set consisting of a plurality of images consecutive in time series, and ground truth data corresponding to each image in the training image set. It is assumed that each image in the training image set contains at least one or more subjects, including at least the subject to be tracked.

[0022] Then, in step S201, the acquiring unit 110 acquires training images and supervised data corresponding to the training images from the training data set 300, and further cuts out an image region of the subject from the training images based on the supervised data. In this embodiment, the acquiring unit 110 acquires three training images, namely, a first training image 311, a second training image 321, and a third training image 331, from the training dataset 300. As described above, the acquiring unit 110 acquires the first training image 311 and the third training image 331 from the same training image set. On the other hand, the acquiring unit 110 acquires the second training image 321 from the same training image set from which the first training image 311 was acquired, or from a different training image set. Note that the acquiring unit 110 may acquire the same image as the first training image 311 as the second training image 321. Furthermore, the acquiring unit 110 acquires, from the training dataset 300, first supervised answer data corresponding to the first training image 311, second supervised answer data corresponding to the second training image 321, and third supervised answer data corresponding to the third training image 331.

[0023] Furthermore, based on the position information of the first correct answer data, acquisition unit 110 specifies the position of the tracked subject from first learning image 311, and cuts out a predetermined image area including the specified position as tracked subject image 312. At this time, the center of tracked subject image 312 cut out by acquisition unit 110 from first learning image 311 is set to the central coordinates of the tracked subject indicated by the first correct answer data, and the width and height of tracked subject image 312 are set to width and height set as predetermined values. Furthermore, based on the position information of the second correct answer data, acquisition unit 110 specifies the position of the tracked subject from second learning image 321, and cuts out an image area of ​​a subject different from the tracked subject at the specified position as non-tracked subject image 322. In other words, acquisition unit 110 cuts out, from among the subjects in second learning image 321, an image area of ​​another subject whose center coordinates are different from the coordinates of the tracked subject indicated in the second correct answer data, as non-tracked subject image 322. Note that the width and height of non-tracked subject image 322 are the same as the width and height of tracked subject image 312. Furthermore, acquisition unit 110 specifies the position of the tracking subject from third learning image 331 based on the position information of the third correct answer data, and cuts out a predetermined image area including the specified position as object tracking subject image 332. At this time, the center of object tracking subject image 332 cut out by acquisition unit 110 from third learning image 331 is set to the central coordinates of the tracking subject indicated by the third correct answer data, and the width and height of object tracking subject image 332 are made the same as the width and height of tracking subject image 312.

[0024] 2, the acquisition unit 110 creates a positive correct likelihood map and a negative correct likelihood map based on the third correct answer data. As described above, the positive correct likelihood map is a map in which the position information of the tracked subject indicated by the third correct answer data is replaced with a likelihood distribution, and the negative correct likelihood map is a map in which all pixel values ​​are 0. In this embodiment, a two-dimensional standard normal distribution expressed by the following equation (1) is used as the likelihood distribution.

[0025]

number

[0026] (μ x ,μ y ) are the central coordinates of the tracked subject indicated by the third correct answer data. Furthermore, a in equation (1) is a parameter that controls the peak value of the distribution and takes a real value greater than 0. Then, the acquisition unit 110 creates a positive correct answer likelihood map by assigning a value of the standard normal distribution expressed by equation (1) to each pixel. Note that in this embodiment, the size of each correct answer likelihood map of the positive correct answer likelihood map and the negative correct answer likelihood map is set to be the same as each subject likelihood map of the positive subject likelihood map and the negative subject likelihood map generated by the subject detection unit 130.

[0027] 2, extraction unit 120 inputs tracked object image 312 to feature amount extraction model 313 to extract tracked object feature amount 314, and similarly inputs non-tracked object image 322 to feature amount extraction model 313 to extract non-tracked object feature amount 324. In this embodiment, feature amount extraction model 313 is made up of a deep neural network with an arbitrary configuration, and the feature amount to be extracted has an arbitrary width and height and at least two or more channels.

[0028] Next, in step S204, subject detection unit 130 obtains a first likelihood and a second likelihood using tracking subject detection model 340. That is, subject detection unit 130 inputs tracking subject feature amount 314 and target tracking subject image 332 into tracking subject detection model 340 to obtain positive subject likelihood map 351, which is the first likelihood. Subject detection unit 130 also inputs non-tracking subject feature amount 324 and target tracking subject image 332 into tracking subject detection model 340 to obtain negative subject likelihood map 361, which is the second likelihood. In this embodiment, tracking subject detection model 340 is configured as a deep neural network, and is a model to which target tracking subject image 332 and tracking subject feature amount 314, or non-tracking subject feature amount 324 and target tracking subject image 332 are input. In addition, tracking subject detection model 340 is also used for calculation processing with feature surfaces of any layer constituting the model. As an example of the computation process with the feature plane of an arbitrary layer, a convolution computation using the subject feature amount as a convolution kernel for the feature plane of a certain layer can be considered.

[0029] 2, loss calculation unit 140 calculates a loss for the object likelihood (positive loss 352) using the positive correct likelihood map and positive object likelihood map 351. Loss calculation unit 140 also calculates a loss for the object likelihood (negative loss 362) using the negative correct likelihood map and negative object likelihood map 361. First, the loss calculation unit 140 calculates a loss (loss) which is a positive loss 352 between the positive ground truth likelihood map and the positive object likelihood map 351 using the cross entropy loss expressed by Equation (2). p Calculate.

[0030]

number

[0031] In equation (2), M p_inf is the positive subject likelihood map, M p_gt is the positive likelihood map, and the sum is calculated for each pixel in the map. p is the weight parameter for the positive loss and can take any real number greater than or equal to 0.

[0032] Similarly, the loss calculation unit 140 calculates a negative loss 362 for the negative ground truth likelihood map and the negative object likelihood map 361 using the cross entropy error expressed by equation (3), n Calculate.

[0033]

number

[0034] In equation (3), M n_inf is the negative subject likelihood map, M n_gt is the negative correct likelihood map, and the sum is calculated for each pixel of the map. n is the weight parameter for the negative loss and can take any real number greater than or equal to 0.

[0035] Then, the loss calculation unit 140 adds the positive loss 352 and the negative loss 362 to calculate the final loss (likelihood loss). For example, if the number of pixels in the map is 1×1 for simplicity, then equation (2) can be expressed as equation (4) below.

[0036]

number

[0037] As can be seen from equation (4), the positive answer likelihood map M p_inf is a positive number according to equation (1), so the positive object likelihood map M n_inf The larger the loss, p In other words, this means that the loss is small when the first likelihood is greater than the second likelihood, and the loss is large when the second likelihood is greater than the first likelihood.

[0038] 2, the weight update unit 150 updates the weight parameters of the feature extraction model 313 and the weight parameters of the tracking object detection model 340 based on the final likelihood loss calculated by the loss calculation unit 140. As a method for updating these weight parameters, for example, backpropagation may be used.

[0039] In the above explanation, an example was given in which one training data set is acquired in one training process, but multiple training data sets may be acquired at once and training processes may be performed. In this case, the processes from step S202 to step S205 are performed for each training data set, and the likelihood loss calculated for each training data set is used to perform the update process in step S206.

[0040] As described above, the learning device 100 of this embodiment not only learns the likelihood of a tracked subject using an image of the tracked subject and the feature quantities of that tracked subject, but also performs learning using the feature quantities of non-tracked subjects. As a result, this embodiment can learn a model that can prevent a non-tracked subject from being output with a high likelihood, which indicates that the subject is a tracked subject. That is, this embodiment can acquire a model that can detect, for example, a tracked subject while filtering out non-tracked subjects. Therefore, this embodiment can prevent erroneous detection of a non-tracked subject as a tracked subject, even when there is a non-tracked subject that is similar in appearance to the tracked subject in terms of shape, color, pattern, etc., and can achieve subject tracking that correctly detects the tracked subject.

[0041] <Second embodiment: learning device> Next, as a second embodiment, an example will be described in which the learning device 100 calculates a distance loss in addition to the likelihood loss described above, and updates the weight parameters of the feature extraction model and the tracking object detection model using the likelihood loss and the distance loss. That is, since learning using the distance loss has the same purpose as learning using the likelihood loss, which is to improve the ability to distinguish between tracked and non-tracked objects, the second embodiment is an example in which a synergistic effect is obtained by performing learning using the distance loss in addition to the likelihood loss. Note that the functional configuration of the learning device 100 of the second embodiment is the same as that shown in FIG. 1, and therefore illustration and description thereof will be omitted.

[0042] Fig. 4 is a flowchart showing the flow of learning processing by the learning device 100 according to the second embodiment. Below, the processing for calculating the distance loss according to this embodiment and the processing for updating the weight parameters of the feature extraction model and the tracking object detection model using the distance loss and likelihood loss will be described with reference to the functional configuration shown in Fig. 1 and the flowchart of Fig. 4. Note that in the flowchart of Fig. 4, the description of the same processing as in the flowchart of Fig. 2 will be omitted.

[0043] In step S401 of FIG. 4 according to the second embodiment, the acquisition unit 110 acquires a fourth training image and corresponding fourth correct answer data from the training dataset in addition to the first to third training images and the first to third correct answer data described above. The fourth training image is an image acquired from the same training image set from which the first training image was acquired, and contains the same tracking object as the first training image. In other words, the tracking object in the fourth training image is the same tracking object as the tracking object in the first training image, but with a time difference. The acquisition unit 110 then identifies the position of the tracking object in the fourth training image based on the position information of the fourth correct answer data, and cuts out a predetermined image area at the identified position. Hereinafter, the image area of ​​the tracking object cut out from the fourth training image will be referred to as a time-difference tracking object image. The center of the time difference tracking subject image that the acquisition unit 110 cuts out from the fourth learning image is set to the central coordinates of the tracking subject indicated by the fourth correct answer data, and the width and height of the time difference tracking subject image are set to be the same as the width and height of the tracking subject image.

[0044] Next, in step S402 of FIG. 4, the acquiring unit 110 creates a positive correct likelihood map and a negative correct likelihood map based on the third correct data, similar to step S202 described above.

[0045] 4, the extraction unit 120 extracts tracking subject feature amounts from the tracking subject image and extracts non-tracking subject feature amounts from the non-tracking subject image, similar to step S203 described above. Furthermore, in this embodiment, the extraction unit 120 also extracts feature amounts from the time-lag tracking subject image cut out from the fourth learning image using the feature extraction model. Hereinafter, the feature amounts extracted from the time-lag tracking subject image will be referred to as time-lag tracking subject feature amounts.

[0046] Next, in step S404 of FIG. 4, object detection unit 130 generates a positive object likelihood map and a negative object likelihood map in the same manner as in step S204 described above. Furthermore, in this embodiment, object detection unit 130 also extracts an object feature vector. In this embodiment, object detection unit 130 extracts the object feature vector from a feature plane (extracted feature plane) of an arbitrary layer constituting the tracking object detection model. However, if the tracking object detection model is configured to perform arithmetic processing between the feature planes of the layers constituting the model and object features, object detection unit 130 selects a layer closer to the output than the layer performing the arithmetic processing as the layer of the extracted feature plane.

[0047] Here, referring to FIG. 5, the process of extracting a subject feature vector will be described, taking as an example a case where a tracked subject image and tracked subject feature quantities are input to a tracked subject detection model. FIG. 5 is a diagram showing, for example, a third training image superimposed with an extracted feature surface enlarged to the same size as the image. The third training image is assumed to be 640 pixels wide x 480 pixels high x 3 channels, and the extracted feature surface is assumed to be 24 pixels wide x 16 pixels high x C channels (C is an arbitrary natural number). Furthermore, the coordinates (x, y) on the third training image are assumed to be in the range of 0≦x≦639, 0≦y≦479, and the coordinates (p, q) on the extracted feature surface are assumed to be in the range of 0≦p≦23, 0≦q≦15.

[0048] The subject detection unit 130 acquires the subject feature vector from a pixel (extracted pixel) on the extracted feature plane that corresponds to the central coordinates of the tracking subject indicated by the third correct answer data. In this embodiment, the subject detection unit 130 performs the following process to determine the extracted pixels: In Fig. 5, it is assumed that the position 501 is the central coordinates (xt, yt) = (312, 276) of the subject to be tracked indicated by the third correct answer data.

[0049] The subject detection unit 130 determines the extracted pixel (pt, qt) by converting the center coordinates of the tracking subject indicated by the third correct answer data into the coordinate scale of the extracted feature plane. For example, the subject detection unit 130 calculates (312×24 / 640)=11.7 for pt and removes the decimal part of this 11.7 to obtain 11, and calculates (276×16 / 480)=9.2 for qt and removes the decimal part of this 9.2 to obtain 9. As a result, the extracted pixel becomes (pt, qt)=(11, 9).

[0050] 5, a rectangle 502 indicates an extracted pixel (pt, qt) corresponding to the center coordinates (xt, yt) of the tracked object on the third training image. The object detection unit 130 acquires the feature amount (width 1 × height 1 × number of channels C) of the extracted pixel (pt, qt) as an object feature vector. The subject feature vector may be a result of calculation using feature vectors of n pixels (n is an arbitrary natural number) surrounding the extracted pixel (pt, qt).

[0051] The subject detection unit 130 also acquires subject feature vectors by the same processing as described above for combinations of subject tracking subject images and non-tracking subject features, and combinations of subject tracking subject images and time-lag tracking subject features. Hereinafter, the subject feature vector acquired using the combination of the tracking subject image and tracking subject feature amounts on the third learning image described above will be referred to as the first subject feature vector. Furthermore, the subject feature vector acquired using the subject tracking subject image and non-tracking subject feature amounts will be referred to as the second subject feature vector, and the subject feature vector acquired using the subject tracking subject image and time-lag tracking subject feature amounts will be referred to as the third subject feature vector.

[0052] 4, the loss calculation unit 140 calculates the likelihood loss in the above-mentioned step S205, and also calculates the loss for the distance between feature amounts (distance loss) based on the first to third object feature vectors. In this embodiment, the loss calculation unit 140 calculates the distance loss using the triplet loss expressed by the following equation (5).

[0053] loss t =max(d1-d2+m,0) Equation (5)

[0054] In equation (5), d1 is the inter-feature distance between the first object feature vector and the third object feature vector, and d2 is the inter-feature distance between the first object feature vector and the second object feature vector. The inter-feature distance is not limited to a specific distance, but the Euclidean distance is used here. Furthermore, m is a margin parameter and takes any real number equal to or greater than 0.

[0055] In this embodiment, the loss function increases as the inter-feature distance d1 increases and d2 decreases. Therefore, it is possible to acquire features such that the inter-feature distance between object feature vectors calculated using object features extracted from two different images is smaller than the inter-feature distance between object feature vectors calculated using tracked object features and non-tracked object features. In other words, this embodiment can acquire losses that increase the inter-feature distance between the first object feature vector and the second object feature vector and decrease the inter-feature distance between the first object feature vector and the third object feature vector. The loss function used in this embodiment is not limited to a specific function as long as it is calculated using two or more of the object feature vectors. For example, contrastive loss may be used as the loss function.

[0056] According to this embodiment, learning is performed so that the distance between feature amounts between subject feature vectors calculated using two subject feature amounts extracted from images between different frames is greater than the distance between feature amounts calculated in a similar manner using tracked subject feature amounts and non-tracked subject feature amounts. This makes it possible to acquire feature amounts that are highly sensitive to differences between tracked subjects and non-tracked subjects, and have low sensitivity to changes in the appearance of the tracked subject between frames, enabling tracking processing that distinguishes between tracked subjects and non-tracked subjects with higher accuracy.

[0057] The learning device 100 of the first and second embodiments described above can also perform learning using videos of multiple different scenes. That is, the learning device 100 of the first and second embodiments can train multiple feature extraction models and tracking object detection models corresponding to multiple different scenes. Examples of multiple video scenes include soccer match scenes, rugby match scenes, horse racing scenes, and motorcycle racing scenes. The feature extraction models and tracking object detection models trained for each of these scenes are models trained using each player in each scene as the tracking target subject. When learning corresponding to such various scenes is performed, the second training images and second ground truth data in the above-described steps S201 and S401 may be acquired from a training dataset separate from the training dataset used to acquire the first training images. This is to prevent the likelihood of identifying a non-tracked subject in a different scene as a tracking subject being acquired.

[0058] <Third embodiment: inference device> Next, as a third embodiment, we will explain an information processing device (hereinafter referred to as an inference device) that performs inference processing for object tracking using a feature extraction model and a tracking object detection model learned by the above-mentioned learning device 100. In this embodiment, an example will be given in which the aforementioned learning device 100 performs learning using videos of multiple different scenes, and multiple trained feature extraction models and tracking subject detection models corresponding to each of the scenes are acquired.

[0059] In the inference device of the third embodiment, a feature extraction model and a tracking object detection model that are suitable for the input video scene are selected from the multiple trained feature extraction models and tracking object detection models, and the inference process is performed using the selected model.

[0060] 6 is a diagram showing an example of the functional configuration of an inference device 600 according to the third embodiment. The inference device 600 selects a model suitable for a scene in a video from among a plurality of trained feature extraction models and tracking subject detection models, and uses the selected feature extraction model and tracking subject detection model to estimate the tracking subject from the video and track the subject. In this embodiment, the estimation and tracking processes of the tracked object are the same as existing processes except that the feature extraction model and the tracked object detection model are models trained by the above-mentioned learning device 100, so detailed explanations thereof will be omitted here. In the third embodiment, the description will focus on the process of selecting a model appropriate for a video scene from a plurality of trained feature extraction models and tracked object detection models trained for each scene by the learning device 100.

[0061] The acquisition unit 610 of the inference device 600 acquires a plurality of consecutive images (video) in time series from a video capture device (not shown) or the like, acquires the most recent image from the video as a first target image, and then cuts out an image area of ​​the tracked subject from the first target image. Note that in the explanation of the inference device 600, the image area of ​​the tracked subject will also be referred to as the tracked subject image. Furthermore, the acquisition unit 610 acquires an image from the video that is at a different time in the time series from the first target image as a second target image. In the case of this embodiment, the second target image is, for example, an image of the frame next to the frame of the first target image.

[0062] In this embodiment, the process of acquiring the latest image from a video as a first target image and cutting out a tracking subject image from the first target image is performed, for example, in response to the following user operation. For example, it is assumed that the display unit 650 of the inference device 600 has a touch panel screen, and that a live view of a video is displayed on the display unit 650. When a user performs an operation on the live view display screen to specify the center position of a desired subject, the acquisition unit 610 acquires the image of the frame at that time as the first target image. Then, the acquisition unit 610 cuts out an image area of ​​a predetermined size centered on the user-specified position from the first target image as the tracking subject image of the first target image. Note that the width and height of the image area cut out from the target image (i.e., the predetermined size) are the same as those used in the process of cutting out the tracking subject image during the learning process described above.

[0063] Extraction unit 620 inputs the tracking subject image cut out from the first target image into a trained feature extraction model to extract features. Note that, also in the description of inference device 600, features extracted from the tracking subject image will be referred to as tracking subject features. Here, the feature extraction model used by extraction unit 620 when extracting features from the tracking subject image of the first target image will be a plurality of feature extraction models trained for each scene by learning device 100. Therefore, the tracking subject features extracted from the tracking subject image of the first target image will be a plurality of tracking subject features respectively extracted by a plurality of feature extraction models trained for each scene.

[0064] The subject detection unit 630 inputs the second target image and the multiple tracking subject features extracted by the extraction unit 620 from the tracking subject image of the first target image into a trained tracking subject detection model, calculates the likelihood of the tracking subject, and obtains a subject likelihood map. Here, the tracking subject detection model to which the tracking subject features are input along with the second target image is a multiple tracking subject detection model trained for each scene by the aforementioned training device 100. However, the trained feature extraction model from which the tracking subject features were extracted and the trained tracking subject detection model to which the tracking subject features are input are models corresponding to the same scene. In other words, for example, the tracking subject features extracted using a feature extraction model trained to correspond to soccer game scenes are input to a tracking subject detection model also trained to correspond to soccer game scenes. In this way, the subject detection unit 630 inputs the tracking subject features extracted from the tracking subject image of the first target image and the second target image into the tracking subject detection models corresponding to each scene, thereby obtaining a multiple subject likelihood map corresponding to each scene.

[0065] The model selection unit 640 selects a pair of feature extraction model and tracking object detection model corresponding to one of the scenes from among the multiple feature extraction models and tracking object detection models trained for each scene by the learning device 100. In the present embodiment, the model selection unit 640 selects the pair of feature extraction model and tracking object detection model corresponding to the object likelihood map with the highest likelihood obtained from the multiple object likelihood maps obtained for each scene by the object detection unit 630. The pair of feature extraction model and tracking object detection model selected by the model selection unit 640 is then used as the model to be used for estimation and object tracking using images of each frame after the second target image frame.

[0066] Fig. 7 is a flowchart showing the processing flow when selecting a feature extraction model and a tracking object detection model in the inference device 600. The processing of the flowchart in Fig. 7 is repeated for each frame of a video. Below, the processing in each functional unit of the inference device 600 according to this embodiment will be described with reference to Figs. 6 and 7. First, in step S701, the acquisition unit 110 acquires a video (a series of consecutive images in time series) from, for example, a video capture device not shown, acquires one image from the acquired video as a target image, and further extracts a tracking subject image from the target image.

[0067] Next, in step S702, inference device 600 branches the process depending on whether the target image acquired in step S701 is the image of the frame when the user specified the center position of the tracking subject, i.e., whether it is the first target image. If inference device 600 determines that it is the first target image, the process proceeds to step S706, but if it is not the first target image, i.e., if it is determined that it is an image of a frame that comes later in time series than the frame of the first target image, the process proceeds to step S703.

[0068] If the process proceeds to step S706, the extraction unit 620 inputs the tracked object image cut out from the first target image into multiple feature extraction models trained to correspond to each scene, and extracts the tracked object features for each of the feature extraction models. After that, the process of the inference device 600 returns to step S701.

[0069] When returning from step S706 to step S701, acquisition unit 110 acquires the image of the next frame in time series from the video as a second target image and cuts out a tracked subject image from that image. At this time, acquisition unit 110 determines the center position of the tracked subject image of the second target image based on the center position of the tracked subject image of the first target image, and cuts out the tracked subject image based on that center position. After that, when proceeding to step S702, inference device 600 determines that it is not the first target image, and processing proceeds to step S703.

[0070] Proceeding to step S703, subject detection unit 630 inputs the tracked subject image cut out from the second target image by acquisition unit 110 and the tracked subject features extracted from the tracked subject image of the previous frame into a plurality of trained tracked subject detection models for each scene. As described above, subject detection unit 630 at this time assumes that the trained feature extraction model that extracted the tracked subject features and the trained tracked subject detection model into which the tracked subject features are input are models corresponding to the same scene. Subject detection unit 630 then outputs a plurality of subject likelihood maps obtained by the plurality of trained tracked subject detection models for each scene to model selection unit 640.

[0071] After step S703, inference device 600 branches the process in step S704 depending on whether the target image acquired in step S701 is the second target image. If inference device 600 determines that it is the second target image, the process proceeds to step S705. On the other hand, if inference device 600 determines that it is not the second target image, that is, that it is an image of a frame that is chronologically later than the frame of the second target image, the process returns to step S701.

[0072] Proceeding to step S705, the model selection unit 640 selects an appropriate model for the video acquired in step S701 from among the multiple feature extraction models and tracking object detection models trained for each scene by the learning device 100. For example, the model selection unit 640 selects a pair of feature extraction model and tracking object detection model for the scene corresponding to the map having the pixel with the highest likelihood from among the object likelihood maps acquired for each scene by the object detection unit 630.

[0073] In this embodiment, an example has been given in which a model is selected based on the highest likelihood among the object likelihood maps corresponding to each scene, but this is not limiting. For example, the model selection unit 640 may select a model for a scene corresponding to the map with the highest total likelihood of all pixel values ​​among the object likelihood maps corresponding to each scene. In this case, since the model for each scene is trained so as to suppress likelihood responses to objects appearing in images of different scenes, an appropriate model can be selected by using the likelihood.

[0074] After step S705, inference device 600 returns the process to step S701 and performs processing on each subsequent frame. That is, the pair of feature extraction model and tracking subject detection model selected by model selection unit 640 is used to estimate the tracking subject using images of each frame subsequent in time series to the frame of the second target image and to perform tracking processing.

[0075] As described above, the inference device 600 according to the third embodiment can select an appropriate model for a scene from among a plurality of feature extraction models and tracking subject detection models learned for each scene by the learning device 100. As a result, the inference device 600 according to this embodiment can perform subject tracking processing using a model with high tracking subject estimation and identification performance according to the scene of the video, enabling tracking with higher accuracy than, for example, when a single model is used.

[0076] <Hardware configuration of information processing device> FIG. 8 is a diagram showing an example of the hardware configuration of an information processing device capable of realizing the learning device 100 and the inference device 600 according to the above-described embodiment. A CPU (Central Processing Unit) 801 controls various devices connected to a bus 808 and executes information processing related to the functional units of the learning device 100 and the inference device 600 described above. A ROM (Read Only Memory) 802 stores a BIOS program and a boot program. A RAM (Random Access Memory) 803 is used as a main storage device for the CPU 801 . An information processing program according to this embodiment is stored in external memory 804. The information processing program stored in external memory 804 is expanded into RAM 803 and executed by CPU 801. This realizes the various functional units of learning device 100 and inference device 600 described above.

[0077] The input unit 805 is a keyboard, mouse, touch panel, or the like, and performs processing related to the input of information from the user. The display unit 806 outputs the processing results of the information processing device to a display device in accordance with instructions from the CPU 801. The function of the display unit 650 described above can be realized by the display unit 806. An I / O (Input / Output) 807 is connected to, for example, an imaging device, an external recording device (not shown), a network, an external display device, and the like, and communicates with them. The bus 808 connects the CPU 801, ROM 802, RAM 803, external memory 804, input unit 805, display unit 806, and I / O 807 in a manner that allows them to communicate with one another.

[0078] The present invention can also be realized by supplying a program that realizes one or more of the functions of the above-described embodiments to a system or device via a network or a storage medium, and having one or more processors in the computer of the system or device read and execute the program.The present invention can also be realized by a circuit (e.g., ASIC) that realizes one or more of the functions. The above-described embodiments are merely examples of specific implementations of the present invention, and the technical scope of the present invention should not be construed as being limited by these embodiments. In other words, the present invention can be implemented in various forms without departing from its technical concept or main features.

[0079] The disclosure of this embodiment includes the following configuration, method, and program. (Configuration 1) an acquisition means for acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction means for acquiring a first feature extracted from the first object in the first image and a second feature extracted from the second object in the second image; a detection means for detecting a first likelihood of the first object in the third image based on the third image and the first feature amount, and detecting a second likelihood of the second object in the third image based on the third image and the second feature amount; a loss calculation means that uses a loss function to obtain a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; an update means for updating an extraction model used by the extraction means to extract the feature amount and a detection model used by the detection means to detect the likelihood so as to reduce the loss acquired by the loss calculation means; An information processing device comprising: (Configuration 2) the acquiring means further generates, based on the first object included in the third image, a first correct likelihood map indicating position information of the first object in the third image as a likelihood distribution, and a second correct likelihood map in which all pixel values ​​indicate the minimum value; the detection means obtains a first object likelihood map in which each pixel indicates a likelihood of being the first object, based on the third image and the first feature amount; and obtains a second object likelihood map in which all pixels indicate a minimum value, based on the third image and the second feature amount; The loss calculation means obtains a first loss based on the first object likelihood map and the first correct likelihood map, obtains a second loss based on the second object likelihood map and the second correct likelihood map, and adds the first loss and the second loss to obtain the loss. 2. The information processing device according to configuration 1, (Configuration 3) the acquisition means further acquires a fourth image including the first subject and having a time difference with respect to the first image; the extraction means further acquires a third feature amount extracted from the first object in the fourth image; the detection means further acquires a first feature vector using the third image and the first feature amount, acquires a second feature vector using the third image and the second feature amount, and acquires a third feature vector using the third image and the third feature amount; the loss calculation means further acquires a loss that increases the inter-feature distance between the first feature vector and the second feature vector and decreases the inter-feature distance between the first feature vector and the third feature vector; The update means updates the extraction model and the detection model based on the loss obtained by the loss calculation means using the likelihood and the loss obtained by the loss calculation means using the feature vector. 3. The information processing device according to configuration 1 or 2. (Configuration 4) the acquiring means acquires the first image, the second image, and the third image from a plurality of time-series consecutive images for a plurality of different scenes, respectively; the extraction means acquires the first feature amount and the second feature amount for each of the scenes; the detection means acquires the first likelihood and the second likelihood for each of the scenes; the loss calculation means acquires, for each of the scenes, a loss that is small when the first likelihood is greater than the second likelihood and that is large when the second likelihood is greater than the first likelihood; The update means updates the extraction model and the detection model for each scene based on the loss acquired for each scene by the loss calculation means. 3. The information processing device according to configuration 1 or 2. (Configuration 5) the acquisition means further acquires a fourth image including the first subject and having a time difference with respect to the first image from a plurality of time-series consecutive images for a plurality of different scenes; the extraction means further acquires a third feature amount extracted from the first subject in the fourth image for each of the scenes; the detection means further acquires, for each of the scenes, a first feature vector using the third image and the first feature amount, a second feature vector using the third image and the second feature amount, and a third feature vector using the third image and the third feature amount; the loss calculation means acquires, for each of the scenes, a loss that increases the inter-feature distance between the first feature vector and the second feature vector and decreases the inter-feature distance between the first feature vector and the third feature vector; The update means updates the extraction model and the detection model for each scene based on the loss acquired by the loss calculation means using the likelihood for each scene and the loss acquired by the loss calculation means using the feature vector. 5. The information processing device according to configuration 4. (Configuration 6) The acquisition means extracts a predetermined image area from the acquired image, the predetermined image area being centered on a central position of the subject; 6. The information processing apparatus according to any one of configurations 1 to 5, wherein the extraction means extracts the feature amount from the cut-out image region. (Configuration 7) 7. The information processing apparatus according to any one of configurations 1 to 6, wherein the first subject and the second subject are similar subjects. (Configuration 8) an acquisition means for acquiring a plurality of images successive in time series; an extraction means for extracting features from a first target image among the plurality of time-series consecutive images acquired by the acquisition means, using an extraction model trained for each of a plurality of different scenes; a detection means for inputting a second target image that is chronologically subsequent to the first target image among the plurality of chronologically consecutive images acquired by the acquisition means, and a feature amount extracted from the first target image by the extraction means using an extraction model trained for each of the scenes, into a detection model trained for each of the scenes, and acquiring a likelihood for the second target image; a model selection means for selecting, based on the likelihood, one of a plurality of extraction models trained for each of the scenes, and for selecting one of a plurality of detection models trained for each of the scenes; and 6. An information processing device, wherein the plurality of extraction models and the detection model trained for each scene are models updated by the information processing device according to configuration 4 or 5. (Configuration 9) The information processing device described in configuration 8, wherein the model selection means selects one extraction model in which the highest likelihood is detected from among the plurality of extraction models trained for each of the scenes, and selects one detection model in which the highest likelihood is detected from among the plurality of detection models trained for each of the scenes. (Configuration 10) The information processing device described in configuration 8, wherein the model selection means selects one extraction model with the highest total likelihood of all pixels from among the plurality of extraction models trained for each scene, and selects one detection model with the highest total likelihood of all pixels from among the plurality of detection models trained for each scene. (Method 1) an acquisition step of acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction step of acquiring a first feature amount extracted from the first object in the first image and a second feature amount extracted from the second object in the second image; a detecting step of detecting a first likelihood in the third image for the first object based on the third image and the first feature amount, and detecting a second likelihood in the third image for the second object based on the third image and the second feature amount; a loss calculation step using a loss function that obtains a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; an updating step of updating an extraction model used to extract the feature amount in the extraction step and a detection model used to detect the likelihood in the detection step so as to reduce the loss acquired in the loss calculation step; An information processing method comprising: (Program 1) A program that causes a computer to function as the information processing device according to any one of configurations 1 to 10. [Explanation of symbols]

[0080] 100: Learning device, 600: Inference device, 110, 610: Acquisition unit, 120, 620: Extraction unit, 130, 630: Object detection unit, 140: Loss calculation unit, 150: Weight update unit, 640: Model selection unit

Claims

1. an acquisition means for acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction means for acquiring a first feature extracted from the first object in the first image and a second feature extracted from the second object in the second image; a detection means for detecting a first likelihood of the first object in the third image based on the third image and the first feature amount, and detecting a second likelihood of the second object in the third image based on the third image and the second feature amount; a loss calculation means that uses a loss function to obtain a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; an update means for updating an extraction model used by the extraction means to extract the feature amount and a detection model used by the detection means to detect the likelihood so as to reduce the loss acquired by the loss calculation means; An information processing device comprising:

2. the acquiring means further generates, based on the first object included in the third image, a first correct likelihood map indicating position information of the first object in the third image as a likelihood distribution, and a second correct likelihood map in which all pixel values ​​indicate the minimum value; the detection means obtains a first object likelihood map in which each pixel indicates a likelihood of being the first object, based on the third image and the first feature amount; and obtains a second object likelihood map in which all pixels indicate a minimum value, based on the third image and the second feature amount; The loss calculation means obtains a first loss based on the first object likelihood map and the first correct likelihood map, obtains a second loss based on the second object likelihood map and the second correct likelihood map, and adds the first loss and the second loss to obtain the loss.

2. The information processing apparatus according to claim 1, wherein:

3. the acquisition means further acquires a fourth image including the first subject and having a time difference with respect to the first image; the extraction means further acquires a third feature amount extracted from the first object in the fourth image; the detection means further acquires a first feature vector using the third image and the first feature amount, acquires a second feature vector using the third image and the second feature amount, and acquires a third feature vector using the third image and the third feature amount; the loss calculation means further acquires a loss that increases the inter-feature distance between the first feature vector and the second feature vector and decreases the inter-feature distance between the first feature vector and the third feature vector; The update means updates the extraction model and the detection model based on the loss obtained by the loss calculation means using the likelihood and the loss obtained by the loss calculation means using the feature vector.

2. The information processing apparatus according to claim 1, wherein:

4. the acquiring means acquires the first image, the second image, and the third image from a plurality of time-series consecutive images for a plurality of different scenes, respectively; the extraction means acquires the first feature amount and the second feature amount for each of the scenes; the detection means acquires the first likelihood and the second likelihood for each of the scenes; the loss calculation means acquires, for each of the scenes, a loss that is small when the first likelihood is greater than the second likelihood and that is large when the second likelihood is greater than the first likelihood; The update means updates the extraction model and the detection model for each scene based on the loss acquired for each scene by the loss calculation means.

2. The information processing apparatus according to claim 1, wherein:

5. the acquisition means further acquires a fourth image including the first subject and having a time difference with respect to the first image from a plurality of time-series consecutive images for a plurality of different scenes; the extraction means further acquires a third feature amount extracted from the first subject in the fourth image for each of the scenes; the detection means further acquires, for each of the scenes, a first feature vector using the third image and the first feature amount, a second feature vector using the third image and the second feature amount, and a third feature vector using the third image and the third feature amount; the loss calculation means acquires, for each of the scenes, a loss that increases the inter-feature distance between the first feature vector and the second feature vector and decreases the inter-feature distance between the first feature vector and the third feature vector; The update means updates the extraction model and the detection model for each scene based on the loss acquired by the loss calculation means using the likelihood for each scene and the loss acquired by the loss calculation means using the feature vector.

5. The information processing apparatus according to claim 4,

6. The acquisition means extracts a predetermined image area from the acquired image, the predetermined image area being centered on a central position of the subject; 6. The information processing apparatus according to claim 1, wherein the extracting means extracts the feature amount from the cut-out image region.

7. 6. The information processing apparatus according to claim 1, wherein the first subject and the second subject are similar subjects.

8. an acquisition means for acquiring a plurality of images successive in time series; an extraction means for extracting features from a first target image among the plurality of time-series consecutive images acquired by the acquisition means, using an extraction model trained for each of a plurality of different scenes; a detection means for inputting a second target image that is chronologically subsequent to the first target image among the plurality of chronologically consecutive images acquired by the acquisition means, and a feature amount extracted from the first target image by the extraction means using an extraction model trained for each of the scenes, into a detection model trained for each of the scenes, and acquiring a likelihood for the second target image; a model selection means for selecting, based on the likelihood, one of a plurality of extraction models trained for each of the scenes, and for selecting one of a plurality of detection models trained for each of the scenes; and The information processing apparatus according to claim 4 , wherein the plurality of extraction models and the plurality of detection models trained for each scene are models updated by the information processing apparatus according to claim 4 .

9. 9. The information processing device according to claim 8, wherein the model selection means selects one extraction model for which the highest likelihood is detected from among the plurality of extraction models trained for each of the scenes, and selects one detection model for which the highest likelihood is detected from among the plurality of detection models trained for each of the scenes.

10. The information processing device according to claim 8, characterized in that the model selection means selects one extraction model with the highest total likelihood of all pixels from among the plurality of extraction models trained for each scene, and selects one detection model with the highest total likelihood of all pixels from among the plurality of detection models trained for each scene.

11. an acquisition step of acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction step of acquiring a first feature amount extracted from the first object in the first image and a second feature amount extracted from the second object in the second image; a detecting step of detecting a first likelihood in the third image for the first object based on the third image and the first feature amount, and detecting a second likelihood in the third image for the second object based on the third image and the second feature amount; a loss calculation step using a loss function that obtains a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; an updating step of updating an extraction model used to extract the feature amount in the extraction step and a detection model used to detect the likelihood in the detection step so as to reduce the loss acquired in the loss calculation step; An information processing method comprising:

12. Computer, an acquisition means for acquiring a first image including a first subject, a second image including a second subject different from the first subject, and a third image including the first subject; an extraction means for acquiring a first feature extracted from the first object in the first image and a second feature extracted from the second object in the second image; a detection means for detecting a first likelihood of the first object in the third image based on the third image and the first feature amount, and detecting a second likelihood of the second object in the third image based on the third image and the second feature amount; a loss calculation means that uses a loss function to obtain a loss such that the loss is small when the first likelihood is greater than the second likelihood and the loss is large when the second likelihood is greater than the first likelihood; an update means for updating an extraction model used by the extraction means to extract the feature amount and a detection model used by the detection means to detect the likelihood so as to reduce the loss acquired by the loss calculation means; A program that causes the device to function as an information processing device having the above.

Images