Research map / Video / Action recognition
Continual 3D Convolutional Neural Networks for Real-Time Processing of Videos
Lukas Hedegaard, Alexandros Iosifidis
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 |
|---|---|---|---|
| convert X3D-S to continual form with unchanged weights | X3D-S · Kinetics-400 | FLOPs: 12.1x fewer FLOPs | Improved |
| convert X3D-M to continual form | X3D-M · Kinetics-400 | FLOPs: 15.1x fewer FLOPs | Improved |
| convert regular 3D CNN with temporal zero-padding to continual form without fine-tuning | X3D-M · Kinetics-400 | top-1 accuracy: 67.24 -> 62.80 | Hurt |
| extend global average pool from 16 to 64 frames | CoX3D-M · Kinetics-400 | top-1 accuracy: 62.80 -> 71.03 | Improved |
| extend global average pool from 13 to 64 frames | CoX3D-S · Kinetics-400 | top-1 accuracy: 60.18 -> 67.33 | Improved |
| extend global average pool from 16 to 64 frames | CoX3D-L · Kinetics-400 | top-1 accuracy: 63.03 -> 71.61 | Improved |
| extend pool size from 8 to 64 frames | CoSlow · Charades | mAP: 21.5 -> 25.2 | Improved |
| extend pool size from 13 to 96 frames (ablation) | CoX3D-S · Kinetics-400 val 15 FPS | top-1 accuracy: 49.2 -> 61.8 | Improved |
| extend pool size from 16 to 96 frames (ablation) | CoX3D-M · Kinetics-400 val 15 FPS | top-1 accuracy: 56.3 -> 67.3 | Improved |
| extend pool size from 16 to 96 frames (ablation) | CoX3D-L · Kinetics-400 val 15 FPS | top-1 accuracy: 53.0 -> 66.3 | Improved |
| extend pool size from 13 to 96 frames (ablation) | CoX3D-S · Kinetics-400 val 15 FPS | FLOPs: 166,565 -> 166,601 | No clear effect |
| initialise state with replicated first-frame features instead of zeros (ablation) | CoX3D-S/M · Kinetics-400 val 15 FPS | transient response: slightly earlier rise | Mixed |
Closest papers by meaning
- Learning Spatio-Temporal Representation With Pseudo-3D Residual Networks (ICCV 2017 · similarity 0.89)
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.89)
- 3D CNNs With Adaptive Temporal Feature Resolutions (CVPR 2021 · similarity 0.87)
- Maximizing Spatio-Temporal Entropy of Deep 3D CNNs for Efficient Video Recognition (ICLR 2023 · similarity 0.87)
- OA-CNNs: Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation (CVPR 2024 · similarity 0.85)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · 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.
