Peking University

Shiliang Zhang

Dongkai Wang

The present application relates to the technical field of deep learning and pose estimation, and more particularly, to a contextual instance decoupling (CID)-based multi-person pose estimation (MPPE) method and apparatus. The method includes: acquiring a preset number of images containing multiple persons; inputting the images containing multiple persons, as a training sample, into a CID-based MPPE model for training; and performing pose estimation on a target image using the trained CID-based MPPE model, the CID-based MPPE model being provided with an instance information abstraction module, a global feature decoupling module and a heatmap estimation module. The method and apparatus of the present application can explore context clues over a greater range, thus being robust to spatial detection errors and superior in both accuracy and efficiency.

Contextual instance decoupling-based multi-person pose estimation method and apparatus

The present application discloses a method and a system for person re-identification, the method including: inputting a training set to a model-to-be-trained, and determining a single-class label and memory features of each image data in the training set; determining multi-class labels through positive label prediction according to the single-class labels and a memory feature set; determining classification scores according to image features of each image data in the training set and the memory feature set; determining a multi-label classification loss according to the multi-class labels and the classification scores; and updating and training the model-to-be-trained to obtain a re-identification model according to the multi-label classification loss. The classification scores are determined according to the image features of each image data in the training set and the memory feature set, which is not affected by the domain gap; the multi-class labels are determined through positive label prediction according to the single-class labels and the memory feature set; then, the multi-label classification loss is determined according to the multi-class labels and the classification scores, and the model-to-be-trained is updated and trained, so that the resulting re-identification model has high performance, strong robustness and low cost.

Growing community of inventors

Beijing, China

Shiliang Zhang