Research map / Video / Video segmentation and tracking
Maximizing Spatio-Temporal Entropy of Deep 3D CNNs for Efficient Video Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Search the 3D architecture with STEntr-Score rather than use the X3D-S baseline architecture | X3D-S-scale model · Something-Something V1-Val | Top1: 44.6 -> 47.1 | Helped |
| Search the 3D architecture with STEntr-Score rather than use the X3D-M baseline architecture | X3D-M-scale model · Something-Something V1-Val | Top1: 47.3 -> 49.4 | Helped |
| Search the 3D architecture with STEntr-Score rather than use the X3D-L baseline architecture | X3D-L-scale model · Something-Something V1-Val | Top1: 49.4 -> 51.1 | Helped |
| Replace the X3D-M architecture with E3D-M | X3D-M · Kinetics400-Val | Top1: 76.0 -> 76.4 | Helped |
| Use STEntr-guided search for a model at the E3D-L scale | E3D-L · Kinetics400-Val | Top1: 77.6 vs. 78.8 | Mixed |
| Use 1×5×5 kernels only in stage 2 and 3×3×3 kernels in other stages (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 44.6 -> 45.15% | Helped |
| Use 1×5×5 kernels only in stage 3 and 3×3×3 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 44.6 -> 44.87% | Helped |
| Use 1×5×5 kernels only in stage 4 and 3×3×3 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 44.6 -> 44.35% | Hurt |
| Use 1×5×5 kernels only in stage 5 and 3×3×3 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 44.6 -> 43.85% | Hurt |
| Use 1×5×5 kernels only in stage 6 and 3×3×3 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 44.6 -> 43.83% | Hurt |
| Use 3×3×3 kernels only in stage 2 and 1×5×5 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 43.43% -> 41.59% | Hurt |
| Use 3×3×3 kernels only in stage 3 and 1×5×5 kernels elsewhere (ablation) | X3D-S kernel selection · Something-Something V1 validation | Top1: 43.43% -> 42.93% | Hurt |
Closest papers by meaning
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.88)
- 2D or not 2D? Adaptive 3D Convolution Selection for Efficient Video Recognition (CVPR 2021 · similarity 0.86)
- Optimization Planning for 3D ConvNets (ICML 2021 · similarity 0.86)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.86)
- Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · similarity 0.84)
- OA-CNNs: Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation (CVPR 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.