Neural network system for 3D pose estimation
A system for 3D pose estimation of people is described, comprising an encoder neural network configured to extract multi-view image features from 2D input images obtained from different camera views, and a decoder neural network configured to receive the multi-view image features and predict 3D joint locations of the people in the 2D input images. The decoder neural network comprising a projective-attention mechanism configured to: determine a 2D projection of a predicted 3D joint location for the camera views and assign them as anchor points; apply an adaptive deformable sampling strategy to gather localized context information of the camera views and to learn deformable offsets, and based on the deformable offsets, determine the deformable points for the anchor points; generate attention weights based on the multi-view image features at the anchor points; and apply the attention weights to aggregate the multi-view image features at the deformable points.
1 . A system comprising one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to implement a neural network system for predicting 3D poses of a plurality of persons in a scene based on 2D input images, each of the 2D input images obtained from a different camera view, the different camera views capturing a visual of the plurality of persons in a scene from different perspectives;
the neural network system comprising an encoder neural network and a decoder neural network, the encoder neural network configured to extract multi-view image features from the 2D input images, the decoder neural network configured to receive the multi-view image features as input, and predict 3D joint locations of each of the plurality of persons;
wherein the decoder neural network comprises decoder layers to regressively refine the predicted 3D joint locations, and wherein each of the decoder layers comprises a projective-attention mechanism, the projective-attention mechanism configured to:
determine a 2D projection of a predicted 3D joint location for each of the different camera views, wherein the predicted 3D joint location is obtained from a preceding decoder layer or is determined by linear methods based on input joint queries;
assign the 2D projection as an anchor point for each of the different camera views;
apply an adaptive deformable sampling strategy to gather localized context information of the different camera views to learn deformable offsets, and based on the deformable offsets, determine the deformable points for each of the anchor points;
generate attention weights based on the multi-view image features at the anchor points; and
apply the attention weights to aggregate the multi-view image features at the deformable points.
2 . The system of claim 1 wherein the input joint queries are joint query embeddings, the joint query embeddings comprising hierarchical embedding of joint queries and person queries that encode a person-joint relation; and wherein each of the decoder layers further comprises a self-attention mechanism configured to perform self-attention on the joint query embeddings.
3 . The system of claim 2 wherein the joint query embeddings are augmented with scene level information that are specific to each of the 2D input images.
4 . The system of claim 1 further comprising a positional encoder, the positional encoder configured to encode the camera ray directions for each of the different camera views into the multi-view image features.
5 . The system of claim 4 wherein the camera ray directions are generated with camera parameters of the different camera views; and wherein encoding the camera ray directions for each of the different camera views into the multi-view image features comprises concatenating channel-wisely the camera ray directions to the corresponding image features, and applying a standard convolution to obtain updated image representations.
6 . The system of claim 5 wherein the projective-attention mechanism is further configured to: receive the updated image representations as input;
generate attention weights based on the updated image representations at the anchor points; and
apply the attention weights to aggregate the updated image representations at the deformable points.
7 . The system of claim 1 wherein each of the decoder layers further comprises a feed forward network block, the feed forward network block configured to apply feed-forward regression to predict the 3D joint locations, and confidence scores of the predicted 3D joint locations.
8 . The system of claim 1 wherein the decoder neural network comprises at least four decoder layers.
9 . The system of claim 1 wherein the encoder neural network comprises a convolutional neural network or a transformer based component.