The authors extend their earlier image-based CycleGRPO method to temporal intervals and timestamped spatial evidence.
Training models to recover regions from descriptions improves visual grounding
Generalized CycleGRPO links description and localization through geometric rewards, improving reported image and video results without reference captions during reinforcement learning.
Top University
Xin Zhang · Haochen Wang · Beibei Lin · Zhaoxiang Zhang · Robby T. Tan · Jason Li · +1 more
NUS · UCAS · NTU
Research Digest··2 min read
Zhang et al.
§