Research map / Video / Action recognition
MorphMLP: An Efficient MLP-Like Backbone for Spatial-Temporal Representation Learning
David Junhao Zhang, Kunchang Li, Yali Wang, Yunpeng Chen, Shashwat Chandra, Yu Qiao, Luoqi Liu, Mike Zheng Shou
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 |
|---|---|---|---|
| expand chunk length progressively (14/28/28/49) instead of constant 14 (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 48.7 -> 50.6 | Improved |
| expand chunk length progressively instead of global-only chunks (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 48.6 -> 50.6 | Improved |
| expand chunk length progressively instead of constant 3 (ablation) | MorphMLP-T · ImageNet-1K (100 epochs) | top-1 accuracy: 78.2 -> 79.6 | Improved |
| expand chunk length progressively (ablation) | MorphMLP-T · ADE20K (ImageNet-pretrained, 100 epochs) | mIoU: 41.0 -> 42.6 | Improved |
| use sequential temporal-then-spatial instead of parallel (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 49.2 -> 50.6 | Improved |
| use skip residual instead of standard residual (T+S) (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 49.8 -> 50.6 | Improved |
| put spatial before temporal with skip residual (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 50.6 -> 31.7 | Hurt |
| replace MorphFC_s with 3x3 conv (ablation) | MorphMLP-T · ImageNet-1K (100 epochs) | top-1 accuracy: 79.6 -> 77.3 | Hurt |
| replace MorphFC_s with 1D group conv of same kernel sizes (ablation) | MorphMLP-T · ImageNet-1K (100 epochs) | top-1 accuracy: 79.6 -> 79.0 | Hurt |
| replace MorphFC_t with 3x1x1 temporal conv (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 50.6 -> 47.9 | Hurt |
| add channel FC to horizontal and vertical FC (ablation) | MorphMLP-S · SSV1 | top-1 accuracy: 49.4 -> 50.6 | Improved |
| add weighted sum after three directions (ablation) | MorphMLP-T · ImageNet-1K (100 epochs) | top-1 accuracy: 79.3 -> 79.6 | Improved |
Closest papers by meaning
- MLP-3D: A MLP-Like 3D Architecture With Grouped Time Mixing (CVPR 2022 · similarity 0.86)
- CycleMLP: A MLP-Like Architecture for Dense Visual Predictions (TPAMI 2023 · similarity 0.86)
- PosMLP-Video: Spatial and Temporal Relative Position Encoding for Efficient Video Recognition (IJCV 2024 · similarity 0.86)
- CycleMLP: A MLP-like Architecture for Dense Prediction (ICLR 2022 · similarity 0.86)
- DistInit: Learning Video Representations Without a Single Labeled Video (ICCV 2019 · similarity 0.85)
- SmallBigNet: Integrating Core and Contextual Views for Video Classification (CVPR 2020 · similarity 0.85)
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.
