Beijing Didi Infinity Technology and Development Co., Ltd.

Didi Research America, LLC

Xun Xu

Ruxiao Bao

Methods and systems of annotation densification for semantic segmentation are disclosed herein. In one example embodiment, such a method includes obtaining image information, obtaining coarse annotation information, performing an image matting operation based upon the image information and based at least indirectly upon the coarse annotation information, and applying an already-trained Convolutional Neural Network (ConvNet) semantic segmentation model in relation to the image information. The method also includes performing a merging operation with respect to both first intermediate information generated at least indirectly by the performing of the image matting operation and second intermediate information generated at least indirectly by the applying of the ConvNet model, where the performing of the merging operation results in fine semantic segmentation annotation information, whereby an additional semantic segmentation model can be trained using that annotation information and the trained additional semantic segmentation model can be applied to generate semantic segmentation output information.

Method and system of annotation densification for semantic segmentation

Methods and systems involving convolutional neural networks as applicable for semantic segmentation, including multi-task convolutional networks employing curriculum based transfer learning, are disclosed herein. In one example embodiment, a method of semantic segmentation involving a convolutional neural network includes training and applying the convolutional neural network. The training of the convolutional neural network includes each of training a semantic segmentation decoder network of the convolutional neural network, generating first feature maps by way of an encoder network of the convolutional neural network, based at least in part upon a dataset received at the encoder network, and training an instance segmentation decoder network of the convolutional neural network based at least in part upon the first feature maps. The applying includes receiving an image, and generating each of a semantic segmentation map and an instance segmentation map in response to the receiving of the image, in a single feedforward pass.

Growing community of inventors

Mountain View, CA, United States of America

Xun Xu