FineRS++: Multi-Hypothesis Reasoning and Segmentation of Small Objects in High-Resolution Images
Published in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) · Under Review · Student First Author, 2026
FineRS++ builds a Qwen2.5-VL-based global-local perception framework with coarse-to-fine MLLM reinforcement learning, GSE + LPR training, and a GRPO localization-aware backtracking reward. On FINERS-4k, it achieves gIoU 55.1, cIoU 46.5, MVQA 83.3%, and OVQA 56.7%.
Recommended citation: YiCheng Yang et al. “FineRS++: Multi-Hypothesis Reasoning and Segmentation of Small Objects in High-Resolution Images.” Under review at TPAMI.