Continuous semantic segmentation device and continuous semantic segmentation model training method thereof

By adopting the methods of knowledge distillation and category separation in continuous semantic segmentation, the problem of confusion between background categories and new categories in new category training is solved, and efficient new category training and model generalization capabilities are achieved.

CN119721042APending Publication Date: 2025-03-28HYUNDAI MOTOR CO LTD +2
0 Cites 0 Cited by

Patent Information

Application Number
CN202411350701.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-08-21
Filing Date
2024-09-26
Publication Date
2025-03-28

AI Technical Summary

Technical Problem

Existing continuous semantic segmentation methods based on knowledge distillation are difficult to effectively train new categories, and the training of new categories is often hindered by the confusion between background categories and new categories.

Method used

By effectively distilling useful knowledge of existing models into new models, using methods of category separation and probability transplantation, ensure that the training of new categories is not confused with old categories and backgrounds. Specific steps include: predicting the output category probability of the current and old model, expanding the category space of the old model, making it consistent with the category space of the current model, calculating the knowledge distillation loss, and updating the model.

Benefits of technology

It realizes that while maintaining the similarity of the model output distribution, it effectively trains new categories, avoids confusion between background categories and new categories, and improves the generalization ability and training efficiency of the model.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention provides a continuous semantic segmentation device based on knowledge distillation and a continuous semantic segmentation model training method thereof. The method for training the continuous semantic segmentation model based on knowledge distillation comprises the following steps: training the continuous semantic segmentation model based on training data; predicting the probability of each category of the output of the current continuous semantic segmentation model generated by training and the probability of each category of the output of the old continuous semantic segmentation model; expanding the category related to the old continuous semantic segmentation model, so that the spatial dimension of the category related to the old continuous semantic segmentation model is the same as the spatial dimension of the category related to the current continuous semantic segmentation model; calculating knowledge distillation loss based on the predicted probability of each category; and updating the current continuous semantic segmentation model based on the knowledge distillation loss.
Need to check novelty before this filing date? Find Prior Art