System and method for learning random-walk label propagation for weakly-supervised semantic segmentation
Systems and methods for training semantic segmentation. Embodiments of the present invention include predicting semantic labeling of each pixel in each of at least one training image using a semantic segmentation model. Further included is predicting semantic boundaries at boundary pixels of objects in the at least one training image using a semantic boundary model concurrently with predicting the semantic labeling. Also included is propagating sparse labels to every pixel in the at least one training image using the predicted semantic boundaries. Additionally, the embodiments include optimizing a loss function according the predicted semantic labeling and the propagated sparse labels to concurrently train the semantic segmentation model and the semantic boundary model to accurately and efficiently generate a learned semantic segmentation model from sparsely annotated training images.
1. A method for training semantic segmentation, comprising:
predicting semantic labeling of each pixel in each of at least one training image using a semantic segmentation model;
predicting semantic boundaries at boundary pixels of objects in the at least one training image using a semantic boundary model concurrently with predicting the semantic labeling;
propagating sparse labels to every pixel in the at least one training image using the predicted semantic boundaries; and
optimizing a loss function according the predicted semantic labeling and the propagated sparse labels to concurrently train the semantic segmentation model and the semantic boundary model to accurately and efficiently generate a learned semantic segmentation model from sparsely annotated training images;
wherein propagating the sparse labels includes a random-walk simulation based on the predicted semantic boundaries and the sparse labels;
wherein the random-walk simulation includes:
determining for each pixel on the training image, a plurality of paths from the pixel to each label of the sparse labels;
determining a probability of each path reaching a particular label; and
determining a probability distribution for each of the sparse labels across each pixel of the training image based on the probability of each path reaching a particular label.
2. The method as recited in claim 1 , wherein each of the plurality of paths includes a 4-connected path.
3. The method as recited in claim 1 , wherein the propagating includes:
generating a partition function summing each probability for each path from the pixel; and
developing a recursion from the partition function based on four nearest neighbor pixels to the pixel for calculating a probability that the pixel has a path reaching the particular label.
4. The method as recited in claim 1 , wherein determining the probability of each path includes decaying the probability exponentially where the path crosses at least one of the predicted semantic boundaries.
5. A system for training a semantic segmentation device, comprising:
a semantic labeling module, configured to predict a semantic labeling of each pixel in each of at least one training image using a semantic segmentation model stored in a memory;
a boundary prediction module configured to predict semantic boundaries at boundary pixels of objects in the at least one training image using a semantic boundary model concurrently with predicting the semantic labeling, wherein the semantic boundary model is stored in a memory;
a label propagation module configured to propagate sparse labels to every pixel in the at least one training image using the predicted semantic boundaries;
an optimization module, including a processor, configured optimize a loss function according the predicted semantic labeling and the propagated sparse labels to concurrently train the semantic labeling module and the boundary prediction module to accurately and efficiently generate a learned semantic segmentation model from sparsely annotated training images;
wherein propagating the sparse labels includes a random-walk simulation based on the predicted semantic boundaries and the sparse labels; and
wherein the random-walk simulation includes:
determining for each pixel on the training image, a plurality of paths from the pixel to each label of the sparse labels;
determining a probability of each path reaching a particular label; and
determining a probability distribution for each of the sparse labels across each pixel of the training image based on the probability of each path reaching a particular label.
6. The system as recited in claim 5 , wherein each of the plurality of paths includes a 4-connected path.
7. The system as recited in claim 5 , wherein the propagating includes:
generating a partition function summing each probability for each path from the pixel; and
developing a recursion from the partition function based on four nearest neighbor pixels to the pixel for calculating a probability that the pixel has a path reaching the particular label before any other label.
8. The system as recited in claim 5 , wherein determining the probability of each path includes decaying the probability exponentially where the path crosses at least one of the predicted semantic boundaries.