Research map / Video / Action recognition
Approximated Bilinear Modules for Temporal Modeling
Xinqi Zhu, Chang Xu, Langwen Hui, Cewu Lu, Dacheng Tao
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 |
|---|---|---|---|
| feed concatenated frames to both ABM entries (ABM-C) instead of adjacent frames separately (ABM-S) (ablation) | ABM-top L=3 · Something-Something v1 val | top-1 accuracy: 30.31 -> 38.32 | Improved |
| stack 3 ABM-C layers instead of 1 (ablation) | ABM-C-top · Something-Something v1 val | top-1 accuracy: 35.49 -> 38.32 | Improved |
| stack 3 ABM-S layers instead of 1 (ablation) | ABM-S-top · Something-Something v1 val | top-1 accuracy: 29.94 -> 30.31 | No clear effect |
| sample snippets of 3 frames instead of single frame per segment (ablation) | ABM-C-top L=3 · Something-Something v1 val | top-1 accuracy: 38.32 -> 42.35 | Improved |
| implant ABMs into CNN layers instead of placing on top (ablation) | ABM-C 8x3 · Something-Something v1 val | top-1 accuracy: 42.35 -> 44.14 | Improved |
| increase dynamic fraction beta from 1/4 to 1/2 to 1 (ablation) | ABM-A-in 16x3 · Something-Something v1 val | top-1 accuracy: 44.89 -> 45.67 -> 46.08 | Improved |
| add shifting inference (ST=3) (ablation) | ABM-A / ABM-C · Something-Something v1 val | top-1 accuracy: 44.89 -> 45.56; 45.67 -> 46.16; 46.08 -> 46.81 | Improved |
| use ABM-C-top instead of TRN as post-CNN temporal model | 8-frame ResNet backbone · Something-Something v1 val | top-1 accuracy: 31.68 -> 38.32 | Improved |
| add snippet sampling and shifting inference (ablation) | ABM-C-in ResNet-50 · Something-Something v1 val | top-1 accuracy: 47.45 -> 49.83 | Improved |
| ensemble ABM-C-in and ABM-A-in | ResNet-34 ABM · Something-Something v1 val | top-1 accuracy: 46.81 -> 49.02 | Improved |
| add optical flow stream with 1:1 weighted fusion | ABM-C-in · Something-Something v1 val | top-1 accuracy: 46.81 -> 50.09 | Improved |
| add ABM-C module to I3D backbone | I3D · Kinetics | top-1 accuracy: 71.1 -> 72.6 | Improved |
Closest papers by meaning
- Temporal Bilinear Networks for Video Action Recognition (AAAI 2019 · similarity 0.91)
- Local Temporal Bilinear Pooling for Fine-Grained Action Parsing (CVPR 2019 · similarity 0.86)
- Temporal Distinct Representation Learning for Action Recognition (ECCV 2020 · similarity 0.86)
- DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification (CVPR 2025 · similarity 0.85)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.85)
- Local-Global Multi-Modal Distillation for Weakly-Supervised Temporal Video Grounding (AAAI 2024 · similarity 0.84)
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.
