Video action quality evaluation method and system based on group-sensitive contrastive regression

Abstract

The invention discloses a video action quality evaluation method and system based on group-sensitive contrastive regression, wherein the method includes: selecting a corresponding example video and example video score according to the current video; Perform spatio-temporal feature extraction and construct merged features; build a group-sensitive regression tree network to regress the merged features to get the final difference score, and combine the final difference score with the sample video score to get the current video score. This method obtains the final target action score by modeling the gap between the target action and the example action, which improves the accuracy of the model's action quality evaluation.

Description

technical field [0001] The invention relates to the technical fields of computer vision and deep learning, in particular to a method and system for evaluating video action quality based on group-sensitive contrastive regression. Background technique [0002] Video Action Quality Assessment (AQA), which aims to evaluate the performance of a specific action, has received increasing attention in recent years because of its crucial role in many real-world applications, including sports and healthcare. important role. Different from conventional action analysis tasks such as action detection and recognition, AQA is more challenging because it needs to predict fine-grained scores from videos containing actions of the same category. Considering the differences between different videos themselves and their action scores, we believe that the key to solving this problem is to discover the differences between videos and predict the scores based on the differences. [0003] In recent ...

AI Technical Summary

Problems solved by technology

Despite some promising results, AQA still faces two challenges: first, since score labels are usually annotated by human judges (e.g., scores for diving competitions are calculated by aggregating scores from different Accurate score prediction is difficult for subjective evaluation; second, the variance between videos used for the AQA task is very small because actors usually perform the same actions in similar environments

Images