Video Mamba Suite: State Space Model as a Versatile Alternative for Video Understanding

Benchmark Model Rank Results
temporal-action-localization-on-activitynet-1-3ActionMamba (InternVideo2-6B)#2mAP: 42.02mAP IOU@0.5: 62.43mAP IOU@0.75: 43.49mAP IOU@0.95: 10.23
temporal-action-localization-on-fineactionActionMamba(InternVideo2-6B)#2mAP: 29.04mAP IOU@0.5: 45.44mAP IOU@0.75: 28.82mAP IOU@0.95: 6.79
temporal-action-localization-on-hacsActionMamba(InternVideo2-6B)#2Average-mAP: 44.56mAP@0.5: 64.02mAP@0.75: 45.71mAP@0.95: 13.34
temporal-action-localization-on-thumos14ActionMamba(InternVideo2-6B)#3Avg mAP (0.3:0.7): 72.72mAP IOU@0.3: 86.89mAP IOU@0.4: 83.09…
dense-video-captioning-on-activitynet-captionsPDVC + DBM#11CIDEr: 26.77
moment-retrieval-on-qvhighlightsvideo-mamba-suite#20mAP: 45.18R@1 IoU=0.5: 66.65R@1 IoU=0.7: 52.19mAP@0.5: 64.37…
zero-shot-video-question-answer-on-egoschema-fullsetTimeMamba#46Accuracy: 38.70
moment-retrieval-on-charades-stavideo-mamba-suite#47R@1 IoU=0.5: 57.18R@1 IoU=0.7: 36.05
action-classification-on-kinetics-400ViViM-S (extra data)#97Acc@1: 80.1