[Publication] Leveraging Motion Priors in Videos for Improving Human Segmentation
Published In ECCV2018. (top-3 computer vision technical conference)
Abstract:
Using weakly-supervised (suedo-label) approach cooperating with domain adaptation to improve human segmentation task performance.
1. Our proposed method improves the performance of human segmentation across multiple scenes and modalities (ie, RGB to Infrared (IR)).
2. We can get additional performance gain, achieved by combining our weakly-supervised active learning approach with domain adaptation approaches.
3. We propose a memory-network-based policy model to select strong candidate segments (referred to as strong motion prior) through reinforcement learning. The selected segments have high precision and are directly used to finetune the model.
[Reference]
Paper: https://openaccess.thecvf.com/content_ECCV_2018/papers/Yu-Ting_Chen_Leveraging_Motion_Priors_ECCV_2018_paper.pdf
Code: https://github.com/Jwy-Leo/LMP