Research map / Video / Video segmentation and tracking
MGMAE: Motion Guided Masking for Video Masked Autoencoding
Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, Limin Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace tube masking with motion guided masking | VideoMAE ViT-B · Something-Something V2 (800e) | Acc@1: 69.6 -> 71.0 | Improved |
| replace tube masking with motion guided masking, longer pre-training | VideoMAE ViT-B · Something-Something V2 (2400e) | Acc@1: 70.8 -> 72.3 | Improved |
| replace tube masking with motion guided masking | VideoMAE ViT-B · Kinetics-400 (1600e) | Acc@1: 81.5 -> 81.8 | Improved |
| replace tube masking with motion guided masking | VideoMAE ViT-B · Kinetics-400 (800e) | Acc@1: 80.0 -> 81.2 | Improved |
| use middle frame as base frame instead of first or random frame (ablation) | MGMAE · SSV2 ablation | Acc@1: 70.5 -> 71.0 | Improved |
| use backward warping instead of forward warping (ablation) | MGMAE · SSV2 ablation | Acc@1: 70.5 -> 71.0 | Improved |
| frame-level top-k sampling instead of clip-level (ablation) | MGMAE · SSV2 ablation | Acc@1: 70.7 -> 71.0 | Improved |
| mixed Gaussian initial mask instead of token or pixel random (ablation) | MGMAE · SSV2 ablation | Acc@1: 70.9 -> 71.0 | Improved |
| tube hole filling instead of random, invisible, visible or previous-map filling (ablation) | MGMAE · SSV2 ablation | Acc@1: 70.6 -> 71.0 | Improved |
| use TVL1 offline flow instead of online RAFT (ablation) | MGMAE · SSV2 ablation | Acc@1: 71.2 vs 71.0 | No clear effect |
| reduce RAFT iterations from 12 to 6 (ablation) | MGMAE · SSV2 ablation | Acc@1: 71.0 -> 71.3 (56 h -> 43 h) | Improved |
| add motion guided masking (costs extra pre-training time) | VideoMAE · SSV2 ablation | pre-training time: 32 h -> 43 h | Hurt |
Later papers that compared with this method
3 later accepted papers report MGMAE as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Self-supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation (NeurIPS 2025 · better on 4 of 4 shared results)
- SIGMA: Sinkhorn-Guided Masked Video Modeling (ECCV 2024 · better on 5 of 5 shared results)
- MART: Masked Affective RepresenTation Learning via Masked Temporal Distribution Distillation (CVPR 2024 · better on 0 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoMAE — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens (ICML 2024 · similarity 0.89)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.88)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · similarity 0.87)
- MotionAura: Generating High-Quality and Motion Consistent Videos using Discrete Diffusion (ICLR 2025 · similarity 0.86)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.86)
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
