Research map / Video / Video segmentation and tracking
SmallBigNet: Integrating Core and Contextual Views for Video Classification
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 temporal convolution with SmallBig 3×1×1 units | ResNet23 · Kinetics400 validation | Top1 accuracy: 68.3 -> 69.0 | Helped |
| Expand the big-view pooling window from 3×1×1 to 3×3×3 (ablation) | SmallBig-R23 · Kinetics400 validation | Top1 accuracy: 69.0 -> 69.5 | Helped |
| Expand the big-view pooling window to 3×7×7 (ablation) | SmallBig-R23 · Kinetics400 validation | Top1 accuracy: 69.0 -> 68.6 | Hurt |
| Progressively convert more residual-block layers into SmallBig units (ablation) | SmallBig-R23 · Kinetics400 validation | Top1 accuracy: 69.5 -> 72.6 | Helped |
| Replace max pooling with average pooling (ablation) | SmallBig-R23 full · Kinetics400 validation | Top1 accuracy: 72.2 vs. 72.6 | Hurt |
| Remove convolution parameter sharing between the views (ablation) | SmallBig-R23 full · Kinetics400 validation | Top1 accuracy: 71.6 vs. 72.6 | Hurt |
| Use a single batch normalization instead of individual BN per view (ablation) | SmallBig-R23 full · Kinetics400 validation | Top1 accuracy: 64.6 vs. 72.6 | Hurt |
| Replace the default extra unit with the simplified extra unit (ablation) | SmallBig-R50 · Kinetics400 validation | Top1 accuracy: 75.8 vs. 76.3 | Hurt |
| Use SmallBig-R23 instead of the 2D R23 baseline | ResNet23 · Kinetics400 validation | Top1 accuracy: 64.1 -> 72.6 | Helped |
| Use SmallBig-R23 instead of NonLocal-R23 | ResNet23 · Kinetics400 validation | Top1 accuracy: 70.2 -> 72.6 | Helped |
| Use SmallBig-R50 instead of TSM-R50 | ResNet50 · Kinetics400 validation | Top1 accuracy: 74.1 -> 76.3 | Helped |
| Use SmallBig-R50 instead of TSM-R50 | ResNet50 · Something-Something V2 validation | Top1 accuracy: 59.1 -> 59.7 | Helped |
Closest papers by meaning
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.89)
- TAda! Temporally-Adaptive Convolutions for Video Understanding (ICLR 2022 · similarity 0.87)
- Convolutional Two-Stream Network Fusion for Video Action Recognition (CVPR 2016 · similarity 0.87)
- 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks (CVPR 2019 · similarity 0.87)
- Compositional Video Understanding with Spatiotemporal Structure-based Transformers (CVPR 2024 · similarity 0.87)
- LSNet: See Large, Focus Small (CVPR 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.