Guo Chen, Yifei HUANG, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, LiMin Wang
Understanding videos is one of the fundamental directions in computer vision research, with extensive efforts dedicated to exploring various architectures such as RNN, 3D CNN, and Transformers. The newly proposed architecture of state space model, e.g., Mamba, shows promising traits to extend its success in long sequence modeling to video modeling. To assess whether Mamba can be a viable alternative to Transformers in the video understanding domain, in this work, we conduct a comprehensive set of studies, probing different roles Mamba can play in modeling videos, while investigating diverse tasks where Mamba could exhibit superiority. We categorize Mamba into four roles for modeling videos, deriving a Video Mamba Suite composed of 14 models/modules, and evaluating them on 12 video understanding tasks. Our extensive experiments reveal the strong potential of Mamba on both video-only and video-language tasks while showing promising efficiency-performance trade-offs. We hope this work could provide valuable data points and insights for future research on video understanding. Code is public: https://github.com/OpenGVLab/video-mamba-suite.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Video | HACS | Average-mAP | 44.56 | ActionMamba(InternVideo2-6B) |
| Video | HACS | mAP@0.5 | 64.02 | ActionMamba(InternVideo2-6B) |
| Video | HACS | mAP@0.75 | 45.71 | ActionMamba(InternVideo2-6B) |
| Video | HACS | mAP@0.95 | 13.34 | ActionMamba(InternVideo2-6B) |
| Video | ActivityNet-1.3 | mAP | 42.02 | ActionMamba (InternVideo2-6B) |
| Video | ActivityNet-1.3 | mAP IOU@0.5 | 62.43 | ActionMamba (InternVideo2-6B) |
| Video | ActivityNet-1.3 | mAP IOU@0.75 | 43.49 | ActionMamba (InternVideo2-6B) |
| Video | ActivityNet-1.3 | mAP IOU@0.95 | 10.23 | ActionMamba (InternVideo2-6B) |
| Video | FineAction | mAP | 29.04 | ActionMamba(InternVideo2-6B) |
| Video | FineAction | mAP IOU@0.5 | 45.44 | ActionMamba(InternVideo2-6B) |
| Video | FineAction | mAP IOU@0.75 | 28.82 | ActionMamba(InternVideo2-6B) |
| Video | FineAction | mAP IOU@0.95 | 6.79 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | Avg mAP (0.3:0.7) | 72.72 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | mAP IOU@0.3 | 86.89 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | mAP IOU@0.4 | 83.09 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | mAP IOU@0.5 | 76.9 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | mAP IOU@0.6 | 65.91 | ActionMamba(InternVideo2-6B) |
| Video | THUMOS’14 | mAP IOU@0.7 | 50.82 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | HACS | Average-mAP | 44.56 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | HACS | mAP@0.5 | 64.02 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | HACS | mAP@0.75 | 45.71 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | HACS | mAP@0.95 | 13.34 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | ActivityNet-1.3 | mAP | 42.02 | ActionMamba (InternVideo2-6B) |
| Temporal Action Localization | ActivityNet-1.3 | mAP IOU@0.5 | 62.43 | ActionMamba (InternVideo2-6B) |
| Temporal Action Localization | ActivityNet-1.3 | mAP IOU@0.75 | 43.49 | ActionMamba (InternVideo2-6B) |
| Temporal Action Localization | ActivityNet-1.3 | mAP IOU@0.95 | 10.23 | ActionMamba (InternVideo2-6B) |
| Temporal Action Localization | FineAction | mAP | 29.04 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | FineAction | mAP IOU@0.5 | 45.44 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | FineAction | mAP IOU@0.75 | 28.82 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | FineAction | mAP IOU@0.95 | 6.79 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | Avg mAP (0.3:0.7) | 72.72 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.3 | 86.89 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.4 | 83.09 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.5 | 76.9 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.6 | 65.91 | ActionMamba(InternVideo2-6B) |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.7 | 50.82 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | HACS | Average-mAP | 44.56 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | HACS | mAP@0.5 | 64.02 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | HACS | mAP@0.75 | 45.71 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | HACS | mAP@0.95 | 13.34 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | ActivityNet-1.3 | mAP | 42.02 | ActionMamba (InternVideo2-6B) |
| Zero-Shot Learning | ActivityNet-1.3 | mAP IOU@0.5 | 62.43 | ActionMamba (InternVideo2-6B) |
| Zero-Shot Learning | ActivityNet-1.3 | mAP IOU@0.75 | 43.49 | ActionMamba (InternVideo2-6B) |
| Zero-Shot Learning | ActivityNet-1.3 | mAP IOU@0.95 | 10.23 | ActionMamba (InternVideo2-6B) |
| Zero-Shot Learning | FineAction | mAP | 29.04 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | FineAction | mAP IOU@0.5 | 45.44 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | FineAction | mAP IOU@0.75 | 28.82 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | FineAction | mAP IOU@0.95 | 6.79 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | Avg mAP (0.3:0.7) | 72.72 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.3 | 86.89 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.4 | 83.09 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.5 | 76.9 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.6 | 65.91 | ActionMamba(InternVideo2-6B) |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.7 | 50.82 | ActionMamba(InternVideo2-6B) |
| Action Localization | HACS | Average-mAP | 44.56 | ActionMamba(InternVideo2-6B) |
| Action Localization | HACS | mAP@0.5 | 64.02 | ActionMamba(InternVideo2-6B) |
| Action Localization | HACS | mAP@0.75 | 45.71 | ActionMamba(InternVideo2-6B) |
| Action Localization | HACS | mAP@0.95 | 13.34 | ActionMamba(InternVideo2-6B) |
| Action Localization | ActivityNet-1.3 | mAP | 42.02 | ActionMamba (InternVideo2-6B) |
| Action Localization | ActivityNet-1.3 | mAP IOU@0.5 | 62.43 | ActionMamba (InternVideo2-6B) |
| Action Localization | ActivityNet-1.3 | mAP IOU@0.75 | 43.49 | ActionMamba (InternVideo2-6B) |
| Action Localization | ActivityNet-1.3 | mAP IOU@0.95 | 10.23 | ActionMamba (InternVideo2-6B) |
| Action Localization | FineAction | mAP | 29.04 | ActionMamba(InternVideo2-6B) |
| Action Localization | FineAction | mAP IOU@0.5 | 45.44 | ActionMamba(InternVideo2-6B) |
| Action Localization | FineAction | mAP IOU@0.75 | 28.82 | ActionMamba(InternVideo2-6B) |
| Action Localization | FineAction | mAP IOU@0.95 | 6.79 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | Avg mAP (0.3:0.7) | 72.72 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | mAP IOU@0.3 | 86.89 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | mAP IOU@0.4 | 83.09 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | mAP IOU@0.5 | 76.9 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | mAP IOU@0.6 | 65.91 | ActionMamba(InternVideo2-6B) |
| Action Localization | THUMOS’14 | mAP IOU@0.7 | 50.82 | ActionMamba(InternVideo2-6B) |
| Moment Retrieval | Charades-STA | R@1 IoU=0.5 | 57.18 | video-mamba-suite |
| Moment Retrieval | Charades-STA | R@1 IoU=0.7 | 36.05 | video-mamba-suite |
| Moment Retrieval | QVHighlights | R@1 IoU=0.5 | 66.65 | video-mamba-suite |
| Moment Retrieval | QVHighlights | R@1 IoU=0.7 | 52.19 | video-mamba-suite |
| Moment Retrieval | QVHighlights | mAP | 45.18 | video-mamba-suite |
| Moment Retrieval | QVHighlights | mAP@0.5 | 64.37 | video-mamba-suite |
| Moment Retrieval | QVHighlights | mAP@0.75 | 46.68 | video-mamba-suite |