Key point detection model training method, key point detection method and device thereof

Abstract

The invention provides a key point detection model training method, a key point detection method and a device thereof, and the training method comprises the steps: obtaining a training sample set; the training samples of the first type comprise images, key point positions and key point availability, and the training samples of the second type comprise images; a teacher network is trained by using the training sample of the first type and the first total loss function; the first total loss function comprises loss for constraining the position and availability of the key point; when the training sample is selected as a second type, the training sample is input into the trained teacher network and student network to obtain a key point position and availability prediction result; the prediction result is input into a second total loss function so as to train a student network through knowledge distillation; and the second total loss function comprises a key point position for constraining the student network to learn the teacher network and the loss of the availability prediction result. Therefore, the marking time consumption can be reduced, the low-quality image detection effect is improved, and the detection real-time performance is improved.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a key point detection model training method, a key point detection method and a device. Background technique [0002] At present, the key point detection task mainly faces the following three problems: [0003] First, the labeling of key point training data is extremely time-consuming, especially when faced with new key point detection tasks, the work of key point labeling is more time-consuming and complicated. Moreover, when different labelers are labeling, they will introduce labeling errors of key points in the labeling task due to personal judgment and labeling differences. The position of the same key point fluctuates in a small range due to labeling errors, which increases the difficulty of key point training convergence. [0004] Second, the detection of low-quality data is difficult. For example, when there are multiple key point detection tasks, when occlusio...

AI Technical Summary

Problems solved by technology

However, in the actual deployment, the inference speed of such a large network is very slow, and the requirements for resources such as memory and video memory are high, which cannot meet the performance requirements in actual use.

Images