Wu, Ye (University of Chinese Academy of Sciences), Wei, Qinggong (University of Chinese Academy of Sciences), Song, Ruiqi (institute of automation,chinese academy of sciences), Cui, Chenglin (University of Chinese Academy of Sciences), Li, Xinqing (University of Chinese Academy of Sciences), Zhu, Fenghua (Institute of Automation, Chinese academy of sciences), Ai, Yunfeng (University of Chinese Academy of Sciences)

A Generative Active Learning Framework for Semantic Segmentation in Autonomous Driving

Scheduled for presentation during the Poster Session "Perception - Semantic segmentation" (ThBT17), Thursday, September 26, 2024, 14:30−16:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Active learning is a machine learning or artificial intelligence method that aims to achieve optimal model performance by actively selecting and labeling the most valuable samples, using the fewest high-quality labeled samples. Relevant works have already been conducted using information uncertainty to select valuable samples. However, its effectiveness for semantic segmentation tasks is relatively poor. In this paper, we propose an active learning framework based on the Generative Adversarial Networks (GANs) architecture for semantic segmentation task, which achieves efficient data selection while simultaneously enhancing semantic segmentation performance. The proposed framework combines the generator (i.e., semantic segmentation network) and the discriminator (i.e., data selection network) for jointly optimizing of both semantic segmentation and data selection tasks. Extensive experiments were conducted on the dataset, which was collected from real-world autonomous driving scenes. Experimental results demonstrate that our method achieved competitive results compared to other SOTA methods.