An object stacking relationship detection method based on a double-layer graph attention network

Abstract

The application discloses a kind of object stacking relationship detection methods based on double-layer graph attention network, belong to computer vision and robot perception technical field;The method comprises: obtaining single-view or multi-view RGB image, after feature extraction, multiple view fusion and target detection, obtain object class and boundary box;Enumerate all ordered object pairs, extract object level features and relationship triple level features;Build the full connection double-layer graph attention network of node and ordered object pair one-to-one correspondence, respectively two kinds of features are input into two subnetworks, update node features by multi-head self-attention;The output features of the same node of two subnetworks are spliced, and the stacking relationship class of each ordered object pair is predicted by a classifier;The application processes object semantic and spatial interaction features in parallel by double-graph attention, effectively captures global dependence, significantly improves the detection accuracy of object superior-inferior relationship in complex stacking scene, and can provide reliable decision basis for robot autonomous grabbing.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and robot perception technology, and in particular to a method for detecting object stacking relationships based on a two-layer graph attention network. Background Technology

[0002] Object stacking relationship detection based on two-layer graph attention networks is an important task in computer vision and robot perception, aiming to identify the hierarchical spatial relationships between objects in a scene (e.g., "object A supports object B" or "object A is above object B"). This technology has wide applications in scenarios such as autonomous robot grasping, warehousing and logistics, and smart homes. Accurate detection of stacking relationships can help robots plan a reasonable grasping sequence, avoiding structural collapse or object damage caused by blind operations.

[0003] In existing technologies, visual stacking relationship detection methods are typically based on single-view RGB images. Object bounding boxes are obtained through target detection, and then appearance features and union region features are extracted for each object pair. Relationship judgment is then performed using feature mapping or graph structures. However, these methods have two main limitations: First, single-view images cannot provide complete features of occluded objects when objects are stacked or severely occluded, causing detection results to rely on the visible parts. This easily leads to simultaneous errors in object recognition and relationship inference. For example, when a box is obscured by a can, it is difficult to determine the supporting relationship between the box and the book below from a single viewpoint. Second, existing methods mostly focus on the local dependencies of relationship triples (object pairs and their union regions), ignoring the global interactions between all object pairs in the scene and failing to model object-level and relationship-level dependencies separately. This local modeling approach makes the model prone to "simultaneous errors in object and relationship" in complex stacking scenes, making it difficult to correct misjudgments using global dependencies and limiting detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide an object stacking relationship detection method based on a two-layer graph attention network. By processing the semantic and spatial interaction features of objects in parallel through two-layer graph attention, it effectively captures global dependencies and significantly improves the detection accuracy of the hierarchical relationship of objects in complex stacking scenarios, which can provide a reliable decision basis for robot autonomous grasping.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for detecting object stacking relationships based on a two-layer graph attention network, comprising the following steps: Acquire single-view or multi-view RGB images of the target scene, extract deep convolutional features from the RGB images of each view, fuse features from multiple views to obtain an enhanced feature map, and perform target detection based on the enhanced feature map to obtain the category and bounding box of all objects in the image. Based on the categories and bounding boxes of all objects, for each pair of different objects, two ordered object pairs are generated in two opposite orders; the object-level features of the first and second objects in each ordered object pair, as well as the triplet-level features of the relationship between the two interaction regions, are extracted. The object-level features and relation triplet-level features are respectively propagated and updated on a fully connected graph structure composed of all ordered object pairs through a multi-head self-attention mechanism, resulting in object-level updated features and relation-level updated features. The object-level update features and relation-level update features belonging to the same ordered object pair are concatenated, and the stacking relation category of each ordered object pair is determined sequentially based on the concatenated features, and the stacking relation detection result is output.

[0006] In some optional embodiments, the extraction methods for the object-level features and relation triplet-level features include: Extract the depth convolution feature vectors of the first object, the second object, and the union region of the two objects from the enhanced feature map. The object-level features are constructed by concatenating the depthwise convolutional feature vectors of the first object and the second object. The first object depth convolution feature vector, the union region depth convolution feature vector, and the second object depth convolution feature vector are concatenated to form the relation triplet-level feature.

[0007] In some optional embodiments, the feature propagation and updating are performed on the fully connected graph structure composed of all ordered object pairs through a multi-head self-attention mechanism, and the specific steps are as follows: An object-level graph attention subnetwork is used to process object-level features, and a relation-level graph attention subnetwork is used to process relation triplet-level features. In each sub-network, multiple independent attention heads are used. Each attention head calculates the attention coefficients between nodes and performs normalization. The features output by each attention head are concatenated as the updated features of the corresponding node. The object-level and relation-level update features are output from the object-level graph attention subnetwork and the relation-level graph attention subnetwork, respectively.

[0008] In some optional embodiments, the multi-view feature fusion involves the following specific steps: When there is only one perspective, use the features of that perspective directly; When multiple viewpoints exist, a two-dimensional entropy is determined based on the features of each viewpoint image, and the viewpoint with the highest entropy value is selected as the main viewpoint. Based on the number of remaining viewpoints, one or two viewpoints with a large structural difference from the main viewpoint are selected as auxiliary viewpoints using the structural similarity index. The global feature vectors of the main viewpoint and auxiliary viewpoints are fused with fixed weights, where the weight of the main viewpoint is greater than the weight of each auxiliary viewpoint, and the sum of the weights of all viewpoints is 1.

[0009] In some optional embodiments, the object detection uses a region proposal network to generate candidate regions. After performing non-maximum suppression to remove redundant boxes on the candidate regions, a preset number of candidate regions with the highest confidence are selected, and the category and bounding box of the selected candidate regions are used as the category and bounding box of all objects in the image.

[0010] In some optional embodiments, the specific steps for determining the stacking relationship category of each ordered pair of objects are as follows: The concatenated features are subjected to three linear transformations in sequence, and ReLU nonlinear activation is performed after each linear transformation. The output dimension of the third linear transformation is equal to the number of stacking relationship categories, resulting in a score vector. The category corresponding to the maximum value in the score vector is then selected as the stacking relationship category of the ordered object pair.

[0011] In some optional embodiments, the method further includes constructing a scene relationship graph based on the stacking relationship categories of each determined ordered pair of objects, and the robot plans the grasping order according to the scene relationship graph and grasps the stacked objects in sequence.

[0012] In some alternative embodiments, the single-view or multi-view RGB images are acquired by a vision sensor mounted on the robot in the working environment.

[0013] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described object stacking relationship detection method based on a two-layer graph attention network.

[0014] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described object stacking relationship detection method based on a two-layer graph attention network.

[0015] The object stacking relationship detection method based on a two-layer graph attention network provided by this invention has at least the following beneficial effects: This invention significantly improves the accuracy of detecting relationships between occluded objects, effectively reducing inference results where both objects and relationships are incorrect simultaneously. It requires no manual feature design, providing end-to-end stacking relationship detection results with only single-view or multi-view RGB images as input, exhibiting excellent generalization ability and real-time performance. Ablation and comparative experiments on publicly available datasets demonstrate that this invention significantly outperforms existing methods in scenarios with a large number of objects and severe occlusion. When deployed in a real robot grasping system, it provides reliable stacking relationship information, assists in planning a safe grasping sequence, effectively avoids structural collapse and object damage, and provides practical decision-making support for autonomous robot operation in unstructured stacked environments. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0017] Figure 1 This is a flowchart of an object stacking relationship detection method based on a two-layer graph attention network according to an embodiment of the present invention; Figure 2 This is a schematic diagram of multi-view feature fusion provided according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an Object-Level Graph Attention Network (O-GAT) according to an embodiment of the present invention; Figure 4 This is an overall framework diagram of the algorithm provided in an embodiment of the present invention; Figure 5 This is a diagram of a robot grasping environment setup according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a robot grasping process in a stacked scenario of five objects according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] One embodiment of the present invention relates to an object stacking relationship detection method based on a two-layer graph attention network. The implementation details of the object stacking relationship detection method based on a two-layer graph attention network in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0020] Step 101: Obtain single-view or multi-view RGB images of the target scene, extract deep convolution features from the RGB images of each view, fuse features from multiple views to obtain an enhanced feature map, and perform target detection based on the enhanced feature map to obtain the category and bounding box of all objects in the image. Specifically, when the input is a single-view RGB image, the image is directly fed into a feature extraction network (such as ResNet-101 or VGG16) to extract global feature maps. When the input is a multi-view RGB image, a primary-secondary view selection strategy based on information entropy is first used to fuse multi-view features.

[0021] First, for each viewpoint image, its two-dimensional entropy is calculated as a quantitative indicator of information content. The formula for calculating the two-dimensional entropy is:

[0022] , ; in, Describes the joint probability distribution. Calculate the frequency of occurrence. The scale of the image.

[0023] Primary and secondary viewpoint selection is based on two-dimensional entropy: All viewpoint images are traversed, and the viewpoint with the highest entropy value is selected as the primary viewpoint. Then, among the remaining viewpoints, the structural similarity index (SSIM) is used to measure the structural similarity between each viewpoint and the primary viewpoint. One or two viewpoints with the largest structural difference from the primary viewpoint (i.e., low SSIM value) and the second highest entropy value are preferentially selected as secondary viewpoints. Specifically: if there is only one remaining viewpoint, that viewpoint is directly used as the sole secondary viewpoint; if there are two or more remaining viewpoints, the two with the largest structural difference and the highest entropy value are selected as secondary viewpoints.

[0024] During feature extraction, the main viewpoint and auxiliary viewpoint images are fed into a feature extraction network with shared weights to obtain their respective global feature vectors. , , (The feature vector corresponding to the missing auxiliary viewpoint is zero). Then, the above feature vectors are linearly weighted and fused with fixed weights:

[0025] When there are three perspectives: ; When there are two perspectives: ; When there is only one perspective: .

[0026] in, , , For the preset fixed weights, satisfy and (If there are only two perspectives, then) and In one specific embodiment, it is possible to take , , (Three-view perspective) or , (Dual perspective).

[0027] Enhanced feature map after fusion The data is fed into the object detector. Object detection employs a Region Proposal Network (RPN) structure. First, a large number of anchor boxes are generated on the enhanced feature map, with each anchor box outputting an object quality score and a bounding box regression offset. After applying regression offsets to the anchor boxes to obtain refined candidate regions, Non-maximum Suppression (NMS) is performed to remove highly overlapping redundant boxes. Then, a predetermined number (e.g., 100) of candidate regions with the highest confidence (object quality score) are selected. The categories and bounding boxes of these candidate regions are used as the final detection results, i.e., the categories and bounding boxes of all objects in the image.

[0028] Multi-view feature fusion diagram as shown Figure 2 As shown in the figure, the two-dimensional entropy of each viewpoint image is calculated, and the viewpoint with the highest entropy value is selected as the main viewpoint. Then, the structural similarity index (SSIM) is used to select one or two views with a large structural difference from the main viewpoint from the remaining views as auxiliary views. The global feature vectors of the main viewpoint and the auxiliary viewpoints are linearly weighted and fused with fixed weights to obtain the enhanced feature map.

[0029] Step 102: Based on the categories and bounding boxes of all objects, generate two ordered object pairs for each pair of different objects in two opposite orders; extract the object-level features of the first and second objects in each ordered object pair, as well as the relationship triplet-level features of the interaction area between the two objects. Let the target detection output a total of n objects, denoted as Each object corresponds to a bounding box. Enumerate all ordered pairs of objects ( ),in common There are 1 pair. In each ordered pair of objects, the first object is... The second object is .

[0030] For each ordered object pair, the following feature vectors are extracted from the original feature map output by the backbone network (i.e., the feature extraction network used in step 101) using the RoI Align operation: The feature vector of the first object According to the bounding box Obtained by cropping and pooling from the feature map; The feature vector of the second object According to the bounding box Obtained by cropping and pooling from the feature map; eigenvectors of the union region : Calculate the union of two bounding boxes The feature map is cropped and pooled based on the union region.

[0031] Will and By concatenating the features, we obtain the object-level feature vector. .Will , , The three are concatenated to obtain the feature vector of the relation triplet. .

[0032] Step 103: The object-level features and relation triplet-level features are propagated and updated on the fully connected graph structure composed of all ordered object pairs through a multi-head self-attention mechanism to obtain object-level updated features and relation-level updated features. Build a containing A fully connected graph network of nodes, where each node uniquely corresponds to an ordered pair of objects. In a graph network, any two nodes are connected by an edge, i.e., a fully connected topology is used.

[0033] Object-Level Graph Attention Subnetwork (O-GAT): The input to this subnetwork is the set of object-level feature vectors for all nodes. For each node Its updated features By aggregating all other nodes ( To remove Information about all nodes (excluding those outside the node) is obtained. Specifically, the node is first calculated. For nodes Attention coefficient:

[0034] ; in, It is a learnable linear transformation matrix. For learnable weight vectors, This represents the vector concatenation operation. Then, the attention coefficients are normalized using the Softmax function to obtain the attention weights.

[0035] ; To enhance the stability of the expression, a multi-head attention mechanism is employed: using An independent attention head ( (The value is a preset positive integer, such as 4 or 8). Each head independently calculates the above attention weights and weighted aggregation, and then the output features of all heads are concatenated: ; in, Representing the The attention weights of the nodes corresponding to each attention head. This is the learnable parameter matrix corresponding to the attention head. The activation function is (e.g., ELU or ReLU). Updated features. It incorporates global dependency information from all other object pairs.

[0036] Relational Level Graph Attention Subnetwork (R-GAT): The input to this subnetwork is the set of relation triplet feature vectors for all nodes. Similar to O-GAT, R-GAT also employs a multi-head self-attention mechanism to update node features on a fully connected graph. First, it computes the node... For nodes Attention coefficient (one-step Softmax normalization):

[0037] ; in Indicates the graph excluding nodes Other nodes, It is a linear transformation weight matrix. This is the weight vector. To further enhance the model's expressive power, R-GAT introduces a multi-head attention mechanism, which calculates the hidden state of a node using K independent attention weights to obtain the node's output features. The calculation formula is shown below:

[0038] ; Updated features It integrates the global dependencies between all relation triples, which can effectively correct relation reasoning errors caused by local occlusion or false detection.

[0039] A schematic diagram of the Object-Level Graph Attention Network (O-GAT) is shown below. Figure 3As shown in the diagram, a fully connected graph network is constructed, where each node corresponds one-to-one with all ordered object pairs. For each node, its input features are object-level features. Through a multi-head attention mechanism, each node aggregates information from all other nodes to update its own features.

[0040] Step 104: Concatenate the object-level update features and relation-level update features belonging to the same ordered object pair, and determine the stacking relation category of each ordered object pair in sequence based on the concatenated features, and output the stacking relation detection result.

[0041] For each node (i.e., each ordered pair of objects), the output features of O-GAT will be... and the output characteristics of R-GAT By splicing the data, the fused features are obtained. Then The input features are fed into the classifier. The classifier consists of three fully connected layers: the first layer maps the input features to an intermediate dimension (e.g., 512) and then to a ReLU activation function; the second layer maps to another intermediate dimension (e.g., 256) and then to a ReLU activation function; the third layer maps to the number of classes. ( The number of stacking relationship categories is preset, such as "supporting", "above", "occluded", "irrelevant", etc. The classifier outputs the score logits for each category, which is then converted into a probability distribution by the Softmax function. The category with the highest probability is taken as the predicted stacking relationship of the ordered object pair.

[0042] The entire model undergoes supervised training using a dataset labeled with object bounding boxes and stacked relationship labels. During training, the object detection part calculates the Intersection over Union (IoU) between the ground truth bounding boxes and candidate regions. Positive and negative samples are assigned based on an IoU threshold (e.g., greater than 0.7 for positive samples, less than 0.3 for negative samples), and detection losses (including classification and regression losses for the RPN and RoI heads) are calculated. The relationship classification part calculates a multi-class cross-entropy loss using the ground truth stacked relationship labels and the prediction results.

[0043] ; in One-hot encoding of the real label. This is the prediction result of the operational relationship. The final total loss is the sum of the detection loss and the relationship classification loss. The parameters of the entire network (including the feature extraction network, object detector, two-layer graph attention network, and classifier) ​​are optimized end-to-end through the backpropagation algorithm.

[0044] During the inference phase (i.e., when the robot is actually running), no actual annotation is required. Steps 101 to 104 above are executed directly to output the stacking relationship category for each ordered pair of objects. The prediction results can be further used to generate a scene relationship graph. Based on this relationship graph (e.g., "object A supports object B, object B supports object C"), the robot plans a safe grasping order (grabbing the topmost object first) to achieve autonomous grasping in complex stacking scenarios.

[0045] The overall framework diagram of the algorithm provided in this embodiment of the invention is as follows: Figure 4 As shown in the diagram, the algorithm framework is divided into three core sub-modules: a feature extraction module, a pairwise object pooling layer, and a two-layer graph neural network module. The input is a single-view or multi-view RGB image, and the output is a graph structure containing the operation relationships between objects.

[0046] Robot grasping environment setup diagram as shown Figure 5 As shown in the figure, an autonomous grasping system for robots, including multi-view cameras, a robotic arm, and grippers, has been constructed. A trained model is deployed to this system, and through real-time inference, the system outputs the stacking relationship of each ordered pair of objects in the scene, generating a relationship graph. The robot then plans the grasping sequence based on this graph.

[0047] A schematic diagram of the robot's grasping process in a scenario where 5 objects are stacked is shown below. Figure 6 As shown in the image, the robot sequentially grabs the topmost object, avoiding the collapse or collision caused by traditional blind grabbing.

[0048] Example 1: This invention was trained and tested on the standard datasets VMRD and REGRAD for robot vision-based operational relationship reasoning. The following are some of the experimental results.

[0049] Table 1. Experiments on multi-view feature fusion and two-layer graph attention network ablation on the REGRAD dataset. Table 2. IA performance on the REGRAD dataset for different numbers of objects. Table 3 Ablation experiments and comparison experiments based on the VMRD dataset (ResNet101) Table 4 Ablation experiments and comparative experiments based on the VMRD dataset (VGG16) Table 5 Performance of Operational Relationship Inference Based on YOLO-V10 Table 6. Success rate of secure data capture in complex scenarios (VGG16) Table 7 Success rate of secure data capture in complex scenarios (ResNet101) Table 8 Performance of the Robotic Gripping System Experimental results show that the method of the present invention can accurately select effective features, filter redundant noise interference, reduce relation reasoning errors of different error types, and significantly improve the robot relation reasoning performance and grasping success rate in complex stacking scenarios.

[0050] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0051] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for detecting object stacking relationships based on a two-layer graph attention network, characterized in that, The method includes: Acquire single-view or multi-view RGB images of the target scene, extract deep convolutional features from the RGB images of each view, fuse features from multiple views to obtain an enhanced feature map, and perform target detection based on the enhanced feature map to obtain the category and bounding box of all objects in the image. Based on the categories and bounding boxes of all objects, for each pair of different objects, two ordered object pairs are generated in two opposite orders; the object-level features of the first and second objects in each ordered object pair, as well as the triplet-level features of the relationship between the two interaction regions, are extracted. The object-level features and relation triplet-level features are respectively propagated and updated on a fully connected graph structure composed of all ordered object pairs through a multi-head self-attention mechanism, resulting in object-level updated features and relation-level updated features. The object-level update features and relation-level update features belonging to the same ordered object pair are concatenated, and the stacking relation category of each ordered object pair is determined sequentially based on the concatenated features, and the stacking relation detection result is output.

2. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The extraction methods for object-level features and relation triplet-level features include: Extract the depth convolution feature vectors of the first object, the second object, and the union region of the two objects from the enhanced feature map. The object-level features are constructed by concatenating the depthwise convolutional feature vectors of the first object and the second object. The first object depth convolution feature vector, the union region depth convolution feature vector, and the second object depth convolution feature vector are concatenated to form the relation triplet-level feature.

3. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The feature propagation and update are performed on the fully connected graph structure composed of all ordered object pairs through a multi-head self-attention mechanism. The specific steps are as follows: An object-level graph attention subnetwork is used to process object-level features, and a relation-level graph attention subnetwork is used to process relation triplet-level features. In each sub-network, multiple independent attention heads are used. Each attention head calculates the attention coefficients between nodes and performs normalization. The features output by each attention head are concatenated as the updated features of the corresponding node. The object-level and relation-level update features are output from the object-level graph attention subnetwork and the relation-level graph attention subnetwork, respectively.

4. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The multi-view feature fusion process involves the following steps: When there is only one perspective, use the features of that perspective directly; When multiple viewpoints exist, a two-dimensional entropy is determined based on the features of each viewpoint image, and the viewpoint with the highest entropy value is selected as the main viewpoint. Based on the number of remaining viewpoints, one or two viewpoints with a large structural difference from the main viewpoint are selected as auxiliary viewpoints using the structural similarity index. The global feature vectors of the main viewpoint and auxiliary viewpoints are fused with fixed weights, where the weight of the main viewpoint is greater than the weight of each auxiliary viewpoint, and the sum of the weights of all viewpoints is 1.

5. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The target detection uses a region proposal network to generate candidate regions. After performing non-maximum suppression to remove redundant boxes on the candidate regions, a preset number of candidate regions with the highest confidence are selected. The category and bounding box of the selected candidate regions are used as the category and bounding box of all objects in the image.

6. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The specific steps for determining the stacking relationship category of each ordered pair of objects are as follows: The concatenated features are subjected to three linear transformations in sequence, and ReLU nonlinear activation is performed after each linear transformation. The output dimension of the third linear transformation is equal to the number of stacking relationship categories, resulting in a score vector. The category corresponding to the maximum value in the score vector is then selected as the stacking relationship category of the ordered object pair.

7. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, It also includes constructing a scene relationship graph by identifying the stacking relationship categories of each ordered pair of objects, and the robot plans the grasping order according to the scene relationship graph and grasps the stacked objects in sequence.

8. The object stacking relationship detection method based on a two-layer graph attention network according to claim 1, characterized in that, The single-view or multi-view RGB images are acquired by the visual sensors mounted on the robot in the working environment.

9. A computer system, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the object stacking relationship detection method based on a two-layer graph attention network as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is capable of performing the object stacking relationship detection method based on a two-layer graph attention network as defined in any one of claims 1 to 8.

Images