Video Mamba Suite: State Space Model as a Versatile Alternative for Video Understanding
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.
PDF AbstractCode
ImageNet
Kinetics
ActivityNet
Kinetics 400
THUMOS14
Charades-STA
YouCook2
Breakfast
EPIC-KITCHENS-100
GTEA
HACS
QVHighlights
FineAction
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Benchmark |
|---|---|---|---|---|---|---|
| Temporal Action Localization | ActivityNet-1.3 | ActionMamba (InternVideo2-6B) | mAP IOU@0.5 | 62.43 | # 1 | |
| mAP | 42.02 | # 1 | ||||
| mAP IOU@0.75 | 43.49 | # 1 | ||||
| mAP IOU@0.95 | 10.23 | # 4 | ||||
| Moment Retrieval | Charades-STA | video-mamba-suite | R@1 IoU=0.5 | 57.18 | # 11 | |
| R@1 IoU=0.7 | 36.05 | # 9 | ||||
| Temporal Action Localization | FineAction | ActionMamba(InternVideo2-6B) | mAP | 29.04 | # 1 | |
| mAP IOU@0.5 | 45.44 | # 1 | ||||
| mAP IOU@0.75 | 28.82 | # 1 | ||||
| mAP IOU@0.95 | 6.79 | # 1 | ||||
| Temporal Action Localization | HACS | ActionMamba(InternVideo2-6B) | Average-mAP | 44.56 | # 1 | |
| mAP@0.5 | 64.02 | # 1 | ||||
| mAP@0.75 | 45.71 | # 1 | ||||
| mAP@0.95 | 13.34 | # 1 | ||||
| Moment Retrieval | QVHighlights | video-mamba-suite | mAP | 45.18 | # 6 | |
| R@1 IoU=0.5 | 66.65 | # 3 | ||||
| R@1 IoU=0.7 | 52.19 | # 2 | ||||
| mAP@0.5 | 64.37 | # 8 | ||||
| mAP@0.75 | 46.68 | # 5 | ||||
| Temporal Action Localization | THUMOS’14 | ActionMamba(InternVideo2-6B) | mAP IOU@0.5 | 76.90 | # 2 | |
| mAP IOU@0.3 | 86.89 | # 2 | ||||
| mAP IOU@0.4 | 83.09 | # 2 | ||||
| mAP IOU@0.6 | 65.91 | # 2 | ||||
| mAP IOU@0.7 | 50.82 | # 2 | ||||
| Avg mAP (0.3:0.7) | 72.72 | # 2 |
