Decomposed Attention Fusion in MLLMs for Training-Free Video Reasoning Segmentation

Open paper
Benchmark Model Rank Results
referring-video-object-segmentation-on-ref-davis17DecAF (training-free, Qwen2.5-VL-7B + SAM2) (extra data)#2J&F: 75.2
referring-video-object-segmentation-on-reasonvosDecAF (training-free, Qwen2.5-VL-7B + SAM2) (extra data)#7J&F: 59.7
referring-video-object-segmentation-on-revosDecAF (training-free, Qwen2.5-VL-7B + SAM2) (extra data)#20J&F: 50.3
referring-video-object-segmentation-on-mevisDecAF (training-free, Qwen2.5-VL-7B + SAM2) (extra data)#28J&F: 43.5
referring-expression-segmentation-on-refer-youtube-vos-2021-public-validationDecAF (training-free, Qwen2.5-VL-7B + SAM2) (extra data)#37J&F: 57.0

Rank counts only results with a code link.