Down-sampling method for point cloud with any sampling size

Abstract

The invention relates to a point cloud downsampling method with any sampling size, which belongs to the field of point cloud data, and comprises the following steps: S1, inputting original point cloud data, and determining a target sampling size range; s2, in the training process, each Epoch dynamically selects the sampling size; s3, generating offset through a point cloud feature extraction module, and generating a final sampling point cloud in combination with an original sampling point obtained by farthest distance sampling FPS; s4, calculating a point-to-point distance value between the sampling point cloud and the original point cloud, and taking the point-to-point distance value as similarity loss; s5, inputting the sampling point cloud into a downstream task network, and calculating task related loss; and S6, repeating the steps S1 to S5 until the model converges. Through one-time training, point clouds of any size can be sampled, limitation of fixed sampling size is avoided, and multi-scale feature learning of the point clouds is realized through different-dimension learning of point cloud features.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud data, and in particular to a point cloud downsampling method of arbitrary sampling size. Background Art

[0002] Point cloud data has become a core data format for fields such as autonomous driving, robotic navigation, industrial inspection, and 3D reconstruction. Point cloud downsampling, a critical preprocessing step, directly impacts the performance and efficiency of subsequent tasks. Appropriate downsampling can reduce computational complexity and improve real-time performance. However, existing downsampling methods still have significant limitations when dealing with dynamic scene requirements (e.g., varying resolution requirements and hardware resource constraints).

[0003] Traditional point cloud downsampling methods primarily involve sampling based on geometric distance (e.g., farthest point sampling (FPS) and random sampling (RS). While computationally efficient, these methods struggle to preserve semantic features, and the sampling results are significantly affected by the initial point distribution. Deep learning-based methods typically train a network model to generate a sampled point cloud or select key point clouds for storage. Network models are typically trained with a fixed sampling size, using a pre-set target sampling size (e.g., 512, 256, or 128) to train a dedicated network. While this can learn task-related features, the model cannot generalize to other sampling sizes, requiring the maintenance of multiple models for different sizes, resulting in a waste of storage and computing resources.

[0004] In recent years, task-driven point cloud downsampling methods (such as classification and reconstruction) have gradually become a research focus. By jointly optimizing the sampling process and downstream tasks, such methods can significantly improve the quality of sampled point clouds. However, existing downsampling methods are limited by a fixed sampling size and cannot flexibly adapt to the dynamic requirements of different tasks and scenarios. Some studies have adopted a radical approach to train models with a sampling size equal to the original point cloud to achieve arbitrary sampling models. Although this approach can improve model adaptability, it increases computational resource consumption and model complexity. Therefore, a more efficient and reasonable point cloud downsampling method is needed that can support sampling point clouds of arbitrary sizes after a single training session and balance the requirements of downstream tasks with computational resource consumption. Summary of the Invention

[0005] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides a point cloud downsampling method with arbitrary sampling size, which has the advantages of being able to support sampling point clouds of arbitrary size after one training and balancing the needs of downstream tasks and the consumption of computing resources. It solves the problem that the existing downsampling method is limited by a fixed sampling size and cannot flexibly adapt to the dynamic needs of different tasks and scenarios.

[0006] (2) Technical solution To achieve the above object, the present invention provides the following technical solution: a point cloud downsampling method of arbitrary sampling size, comprising the following steps: Step S1, input the original point cloud data and determine the target sampling size range; Step S2: During the training process, each epoch dynamically selects the sampling size; Step S3: Generate an offset through the point cloud feature extraction module and combine it with the original sampling points obtained by the farthest sampling FPS to generate the final sampling point cloud; Step S4, calculating the point-to-point distance between the sampled point cloud and the original point cloud as the similarity loss; Step S5: input the sampled point cloud into the downstream task network and calculate the task-related loss; Step S6, repeat S1 to S5 until the model converges.

[0007] Preferably, in step S1, the target sampling size range in this method can be set arbitrarily, and the target sampling size range is set to [N / 16, N / 8, N / 4, N / 2], where N is the total number of points in the original point cloud. In the above dataset, N=1024, that is, the sampling range is [64,128,256,512]. For each sampling size, the corresponding loss value is calculated during the training process. The loss value is used to measure the geometric similarity between the sampled point cloud and the original point cloud, as well as the effect in the task network.

[0008] Preferably, in step S2, in each training epoch, the loss value distribution is calculated according to the preset sampling size range to dynamically select the sampling size of the current epoch. The selection is based on the loss value of each sampling size, and the sampling size with a larger loss value is more likely to be selected.

[0009] Preferably, in step S3, the offset is generated by the point cloud feature extraction module, and the original sampling points obtained by the farthest distance sampling (FPS) are combined to generate the final sampling point cloud, as follows: S3.1 Input point cloud data: The original point cloud data, with the shape of [B, 3, N], where B represents the batch size, 3 represents the 3D coordinates of each point, and N represents the number of points in the point cloud, is fed into the network as input; S3.2 Local feature extraction: S3.21 uses the KNN algorithm to select k nearest neighbor points for each point. The coordinates and features of these neighbor points will be extracted for subsequent feature calculations. S3.22 For each point and its neighbors, calculate the coordinate difference, the 3D coordinate difference of each point + the 3D coordinate itself, to obtain the coordinates of the neighboring points of the shape [B,6,N,k], where 6 represents the 6-dimensional features of each point and its neighbors, 3D coordinate difference + 3D coordinate; S3.23 performs convolution on the features of each point and its neighbors to obtain the features of the neighboring points, with the shape of [B, 2C, N, k], where C is the feature dimension; S3.24 uses a pooling operation to concatenate point-wise features and dimensional features to generate a local context encoding for each point; S3.3 Global Feature Extraction: We perform mean pooling on both the N dimension (the number of points in the point cloud) and the C dimension (the feature dimension of each point) of the point cloud to extract global features. This step captures the overall structural information of the point cloud and provides global context for subsequent downstream tasks. S3.4 Feature Fusion: Perform bilinear fusion on local and global features. This is done by combining local and global features through matrix multiplication. The fused feature representation is richer and more representative, and is ultimately converted into offset values using linear constraints. S3.5 Generate sampling point cloud: S3.51 uses the Farthest Point Sampling (FPS) algorithm to select some points in the original point cloud as the initial sampling point cloud based on the distance between points; S3.52 The initial sampling points are combined with the offset values obtained through deep learning to generate the final sampling point cloud [B,k,3]; After feature extraction and sampling point cloud generation, the final downsampled point cloud is obtained; assuming that the shape of the downsampled point cloud is [B, K, 3], where B represents the batch size, 3 represents the three-dimensional coordinates of each sampling point, and k represents the number of points in the downsampled point cloud.

[0010] Preferably, in step S4, the similarity loss is calculated in the following manner:

[0011] in is the sampling point cloud, is the original point cloud, , are the points in the original point cloud and the sampled point cloud set, Represents the square of the Euclidean distance between two points; by calculating the point-to-point distance between the sampled point cloud and the original point cloud, the similarity between the sampled point cloud and the original point cloud is evaluated to ensure that the sampled point cloud retains the shape and structural characteristics of the original point cloud as much as possible; Among them, the point cloud feature extraction network includes the following modules: S4.1 Local feature extraction module: selects nearby points through KNN and uses convolutional layers to extract local features; S4.2 Global feature extraction module: perform mean pooling on the number of point clouds in the N dimension and the feature dimension of the point cloud in the C dimension to obtain the global features of the point cloud; S4.3 Feature fusion module: fuses global features with local features to preserve the geometric structure and semantic information of the point cloud.

[0012] Preferably, in step S5, S5.1. In classification tasks, the task network PointNet outputs a class probability distribution. A cross-entropy loss function is typically used to calculate the difference between the predicted class and the true label. Through the cross-entropy loss function, the model learns how to accurately classify the sampled point cloud. S5.2. In the reconstruction task, the task network PointAE outputs a reconstructed point cloud. The reconstructed point cloud has the same scale as the original point cloud. The similarity between the original and reconstructed point clouds is measured by calculating the point-to-point distance between the two sets of point clouds. A smaller distance indicates a more accurate reconstruction. Assume that the above epoch selects 64 points as the sampling size, calculates the loss and updates the network; the loss values are also calculated for the sampling range [128, 256, 512], and the loss values of each sampling size are stored in the loss value list of the corresponding size for subsequent updating of the selection probability; After every M rounds of training, which is set to 10 rounds in this invention, the selection probability of each sampling size is recalculated; the specific formula is as follows:

[0013] in represents the probability of the i-th sampling size, the sampling size list [64, 128, 256, 512]) being selected, represents the exponential operation, which amplifies the probability of high loss sampling size by exponential operation while keeping the distribution smooth. It represents the average loss of the i-th sampling size within these 10 epochs.

[0014] Preferably, the task-related loss is determined according to the type of specific downstream task. In the classification task, the cross entropy loss function can be used; in the reconstruction task, the chamfer distance can be used as the loss function; in the present invention, PointNet is selected as the classification task network and PointAE is selected as the reconstruction task network. Before training the sampling network, the task network is first pre-trained, and the parameters of the task network are frozen after the training is completed to ensure its stability.

[0015] Preferably, the model balances the quality of downsampled point clouds and subsequent task performance by learning similarity loss and task loss.

[0016] (3) Beneficial effects Compared with existing technologies, this paper provides a point cloud downsampling method with arbitrary sampling size. By using network-based offset prediction and fine-tuning of the farthest point sampling, and jointly training with different task networks, it achieves a balance between geometric fidelity, task adaptability, and computational efficiency. Its core advantages are: Multi-scale feature extraction: The KNN algorithm is used to divide the local neighborhood of each point and extract local geometric features. It also integrates the spatial coordinate information (N-dimension) and additional attribute information (C-dimension) of the point cloud data to achieve multi-scale feature extraction of the point cloud. Arbitrary sampling size: During training, the sampling size to be optimized is determined based on the loss, and different sampling sizes are optimized. Initial sampling points are first obtained through FPS. The positions of these initial points are then fine-tuned using the offsets learned by the network from global features to obtain the final downsampling results. This mechanism avoids the resource waste caused by repeated training and deployment of multiple models for different sampling sizes, thereby improving the model's versatility. End-to-end optimization: Geometric consistency is constrained through a similarity loss, and the quality of the downsampled point cloud is evaluated and optimized in conjunction with the downstream task network. This overcomes the shortcoming of the FPS method, which is independent of subsequent tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a point cloud downsampling method with arbitrary sampling size proposed by the present invention; Figure 2 Schematic diagram of a point cloud downsampling method of arbitrary sampling size and a point cloud downsampling network model of arbitrary sampling size proposed by the present invention; Figure 3 This is a schematic diagram of the distribution of the sampled point cloud in the original point cloud when the sampling size is 32 in the point cloud downsampling method of arbitrary sampling size proposed by the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] The present invention provides a point cloud downsampling method of arbitrary sampling size, which involves the field of point cloud downsampling. Figures 1 to 3 , including the following steps: Step S1: input the original point cloud data and determine the target sampling size range; S1.1 uses the ModelNet40 dataset for classification tasks and the ShapeNet Core55 dataset for reconstruction tasks; S1.2 Determine the sampling size range, which is [N / 16, N / 8, N / 4, N / 2]. In the above dataset, N=1024, that is, the sampling range is [64, 128, 256, 512]; S1.3 For each sampling size, the corresponding loss value is calculated during training. This loss value is used to measure the geometric similarity between the sampled point cloud and the original point cloud, as well as the effect in the task network.

[0020] Step S2: During the training process, each epoch dynamically selects the sampling size; Based on the loss distribution calculated in step 1, the sampling size for the current epoch is selected probabilistically. Sampling sizes with larger loss values have a higher probability of selection, ensuring that the model focuses more on optimizing poorly sampled point clouds. Initially, all sampling sizes are selected with equal probability.

[0021] Step S3: Generate offsets through the point cloud feature extraction module and combine them with the original sampling points obtained by the farthest distance sampling (FPS) to generate the final sampling point cloud. The details are as follows: S3.1 Input point cloud data: The original point cloud data (shape is [B,3,N], where B represents the batch size, 3 represents the three-dimensional coordinates of each point, and N represents the number of points in the point cloud) is fed into the network as input.

[0022] S3.2 Local feature extraction: S3.21 uses the KNN algorithm to select the k nearest neighbor points for each point. The coordinates and features of these neighbor points are extracted and used for subsequent feature calculations.

[0023] S3.22 For each point and its neighbors, calculate the coordinate difference (the 3D coordinate difference of each point + the 3D coordinate itself) to obtain the coordinates of the neighboring points with the shape of [B,6,N,k], where 6 represents the 6-dimensional features of each point and its neighbors (3D coordinate difference + 3D coordinate).

[0024] S3.23 performs convolution on the features of each point and its neighbors to obtain the features of the neighboring points, with a shape of [B, 2C, N, k], where C is the feature dimension.

[0025] S3.24 uses pooling operations to concatenate point-wise features and dimensional features to generate local context encoding for each point.

[0026] S3.3 Global Feature Extraction: We perform mean pooling on both the N-dimension (the number of points in the point cloud) and the C-dimension (the feature dimension of each point) of the point cloud to extract global features. This step helps capture the overall structural information of the point cloud and provides global context for subsequent downstream tasks.

[0027] S3.4 Feature fusion: Perform bilinear fusion on local features and global features. Local features and global features are combined through matrix multiplication. The fused feature representation will be richer and more representative, and finally converted into offset values through linear constraints.

[0028] S3.5 Generate sampling point cloud: S3.51 uses the Farthest Point Sampling (FPS) algorithm to select some points in the original point cloud as the initial sampling point cloud based on the distance between points.

[0029] S3.52 The initial sampling points are combined with the offset values obtained through deep learning to generate the final sampling point cloud [B,k,3]; After feature extraction and sampling point cloud generation, the final downsampled point cloud is obtained. Assume that the shape of the downsampled point cloud is [B, K, 3], where B represents the batch size, 3 represents the 3D coordinates of each sampling point, and k represents the number of points in the downsampled point cloud.

[0030] Step S4: By calculating the point-to-point distance between the sampled point cloud and the original point cloud, the similarity between the sampled point cloud and the original point cloud is evaluated to ensure that the sampled point cloud retains the shape and structural features of the original point cloud as much as possible. The specific formula is as follows:

[0031] in is the sampling point cloud, is the original point cloud, , are the points in the original point cloud and the sampled point cloud set, Represents the square of the Euclidean distance between two points.

[0032] Among them, the point cloud feature extraction network includes the following modules: S4.1 Local feature extraction module: selects nearby points through KNN and uses convolutional layers to extract local features; S4.2 Global feature extraction module: perform mean pooling on the number of point clouds in the N dimension and the feature dimension of the point cloud in the C dimension to obtain the global features of the point cloud; S4.3 Feature fusion module: fuses global features with local features to preserve the geometric structure and semantic information of the point cloud.

[0033] Step S5: In addition to the above similarity loss, the task network provides a corresponding loss to evaluate the quality of the sampled point cloud: S5.1 In classification tasks, the task network (PointNet) outputs a class probability distribution. A cross-entropy loss function is typically used to calculate the difference between the predicted class and the true label. Through the cross-entropy loss function, the model learns how to accurately classify the sampled point cloud.

[0034] In the reconstruction task, the task network (PointAE) outputs a reconstructed point cloud. The reconstructed point cloud has the same scale as the original point cloud. The point-to-point distance between the original and reconstructed point clouds is calculated to measure the similarity between the two sets of point clouds. A smaller distance indicates a more accurate reconstruction.

[0035] Assuming that the sampling size in the above epoch is 64 points, the loss is calculated and the network is updated. Losses are also calculated for the sampling range [128, 256, 512]. The loss values for each sampling size are stored in the loss value list of the corresponding size to facilitate subsequent updates to the selection probability.

[0036] After every M rounds of training (e.g., 5 or 10 rounds, set to 10 rounds in this invention), the selection probability of each sampling size is recalculated. The specific formula is as follows:

[0037] in represents the probability that the i-th sampling size (sampling size list [64, 128, 256, 512]) is selected, represents the exponential operation, which amplifies the probability of high loss sampling size by exponential operation while keeping the distribution smooth. It represents the average loss of the i-th sampling size within these 10 epochs.

[0038] Step S6: repeat steps S1 to S5 until the model converges.

[0039] Experimental setup: The experimental operating system of this invention is Ubuntu20.04, the CPU is Intel 12th i5-12400, the GPU is Nvidia GeForce GTX 3090, and the development environment used is Python 3.7, PyTorch 1.10.1, and CUDA11.3.

[0040] To verify the effectiveness and practicality of our proposed point cloud downsampling method, we tested it on multiple public datasets. We used the task network frameworks PointNet (for classification) and PointAE (for reconstruction), trained the model at different sampling sizes, and calculated the task losses.

[0041] In order to evaluate the impact of downsampling on task performance, we tested the performance of classification tasks and reconstruction tasks respectively. As shown in Table 1 and Table 2, the classification accuracy is higher and the reconstruction error value is smaller. Figure 3 The distribution of the sampled point cloud in the original point cloud is shown in Figure 2. It can be seen that the sampled point cloud is discretely and evenly distributed around the original point cloud.

[0042] Experimental results on classification and reconstruction tasks demonstrate that, compared to traditional methods, our dynamic sampling size selection strategy enables the model to adapt to varying numbers of sampling points, thereby avoiding the generalization issues and performance degradation associated with fixed sampling sizes in traditional methods. By handling different sampling sizes in a single training session, the time and resource consumption associated with repeated training for varying sampling sizes is reduced.

[0043] Table 1 Classification task accuracy in the ModelNet40 dataset Table 2 ShapeNet Core55 dataset reconstruction task error values While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A point cloud downsampling method of arbitrary sampling size, characterized in that: The following steps are involved: Step S1, input the original point cloud data and determine the target sampling size range; Step S2: During the training process, each epoch dynamically selects the sampling size; Step S3: Generate an offset through the point cloud feature extraction module and combine it with the original sampling points obtained by the farthest sampling FPS to generate the final sampling point cloud; Step S4, calculating the point-to-point distance between the sampled point cloud and the original point cloud as the similarity loss; Step S5: input the sampled point cloud into the downstream task network and calculate the task-related loss; Step S6: repeat steps S1 to S5 until the model converges.

2. A point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: In step S1, the target sampling size range in this method can be set arbitrarily. The target sampling size range is set to [N / 16, N / 8, N / 4, N / 2], where N is the total number of points in the original point cloud. In the above dataset, N=1024, that is, the sampling range is [64, 128, 256, 512]. For each sampling size, the corresponding loss value is calculated during the training process. The loss value is used to measure the geometric similarity between the sampled point cloud and the original point cloud.

3. The point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: In step S2, in each training epoch, the loss value distribution is calculated according to the preset sampling size range to dynamically select the sampling size of the current epoch. The selection is based on the loss value of each sampling size. The sampling size with a larger loss value is more likely to be selected.

4. The point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: In step S3, the offset is generated by the point cloud feature extraction module, and combined with the original sampling points obtained by the farthest distance sampling FPS, the final sampling point cloud is generated, as follows: S3.1 Input point cloud data: The original point cloud data, with the shape of [B, 3, N], where B represents the batch size, 3 represents the 3D coordinates of each point, and N represents the number of points in the point cloud, is fed into the network as input; S3.2 Local feature extraction: S3.21 uses the KNN algorithm to select k nearest neighbor points for each point. The coordinates and features of the neighbor points will be extracted for subsequent feature calculations: S3.22 For each point and its neighbors, calculate the coordinate difference, the 3D coordinate difference of each point + the 3D coordinate itself, to obtain the coordinates of the neighboring points of the shape [B,6,N,k], where 6 represents the 6-dimensional features of each point and its neighbors, 3D coordinate difference + 3D coordinate; S3.23 performs convolution on the features of each point and its neighbors to obtain the features of the neighboring points, with the shape of [B, 2C, N, k], where C is the feature dimension; S3.24 uses a pooling operation to concatenate point-wise features and dimensional features to generate a local context encoding for each point; S3.3 Global Feature Extraction: Perform mean pooling on the N dimension (the number of points in the point cloud) and the C dimension (the feature dimension of each point) of the point cloud to extract the global features of the point cloud. This step is used to capture the overall structural information of the point cloud and provide global context information for subsequent downstream tasks; S3.4 Feature Fusion: Perform bilinear fusion on local and global features. This is done by combining local and global features through matrix multiplication. The fused feature representation is richer and more representative, and is ultimately converted into offset values using linear constraints. S3.5 Generate sampling point cloud: S3.51 uses the Farthest Point Sampling (FPS) algorithm to select some points in the original point cloud as the initial sampling point cloud based on the distance between points; S3.52 The initial sampling points are combined with the offset values obtained through deep learning to generate the final sampling point cloud [B,k,3]; After feature extraction and sampling point cloud generation, the final downsampled point cloud is obtained; assuming that the shape of the downsampled point cloud is [B, K, 3], where B represents the batch size, 3 represents the three-dimensional coordinates of each sampling point, and k represents the number of points in the downsampled point cloud.

5. The point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: In step S4, the similarity loss is calculated as follows:

6. Among them is the sampling point cloud, is the original point cloud, , are the points in the original point cloud and the sampled point cloud set, Represents the square of the Euclidean distance between two points; by calculating the point-to-point distance between the sampled point cloud and the original point cloud, the similarity between the sampled point cloud and the original point cloud is evaluated to ensure that the sampled point cloud retains the shape and structural characteristics of the original point cloud as much as possible; Among them, the point cloud feature extraction network includes the following modules: S4.1 Local feature extraction module: selects nearby points through KNN and uses convolutional layers to extract local features; S4.

2. Global feature extraction module: Perform mean pooling on the number of point clouds in the N-dimensional point cloud and the feature dimension of the point cloud in the C-dimensional point cloud to obtain the global features of the point cloud; S4.

3. Feature fusion module: fuses global features with local features to preserve the geometric structure and semantic information of the point cloud.

7. The point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: In step 5, S5.1 In classification tasks, the task network PointNet outputs a category probability distribution; a cross-entropy loss function is usually used to calculate the difference between the predicted category and the true label; through the cross-entropy loss function, the model learns how to accurately classify the sampled point cloud; In the reconstruction task, the task network PointAE outputs a reconstructed point cloud. The reconstructed point cloud has the same scale as the original point cloud. The point-to-point distance between the original and reconstructed point clouds is calculated to measure the similarity between the two groups of point clouds. A smaller distance indicates a more accurate reconstruction. Assume that the above epoch selects 64 points as the sampling size, calculates the loss and updates the network; also calculates the loss value for the sampling range [128, 256, 512], and stores the loss value of each sampling size in the loss value list of the corresponding size, which is used for subsequent update selection probability. After every M rounds of training, which is set to 10 rounds in this invention, the selection probability of each sampling size is recalculated; the specific formula is as follows:

8. Among them represents the probability of the i-th sampling size, the sampling size list [64, 128, 256, 512]) being selected, represents the exponential operation, which amplifies the probability of high loss sampling size by exponential operation while keeping the distribution smooth. It represents the average loss of the i-th sampling size within these 10 epochs.

9. A point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: The task-related loss is determined according to the type of specific downstream task. In the classification task, the cross entropy loss function is used; in the reconstruction task, the chamfer distance can be used as the loss function; in the present invention, PointNet is selected as the classification task network and PointAE is selected as the reconstruction task network. Before training the sampling network, the task network is first pre-trained, and the parameters of the task network are frozen after the training is completed to ensure its stability.

10. A point cloud downsampling method of arbitrary sampling size according to claim 1, characterized in that: The model balances the quality of downsampled point clouds and subsequent task performance by learning similarity loss and task loss.

Images