Research map / Video / Action recognition
3DV: 3D Dynamic Voxel for Action Recognition in Depth Video
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 |
|---|---|---|---|
| Encode motion values in addition to voxel coordinates (ablation) | 3DV-PointNet++ · NTU RGB+D 120 | accuracy (%): 61.4 -> 75.1; 68.9 -> 87.4 | Helped |
| Add temporal splits to the global motion representation (ablation) | 3DV-PointNet++ · NTU RGB+D 120 | accuracy (%): 75.1 -> 75.8 -> 76.9; 87.4 -> 89.6 -> 92.5 | Helped |
| Add an appearance stream to the 3DV motion stream (ablation) | 3DV-PointNet++ · NTU RGB+D 120 and NTU RGB+D 60 | accuracy (%): 76.9 -> 82.4; 92.5 -> 93.5; 84.5 -> 88.8; 95.4 -> 96.3 | Helped |
| Use appearance alone instead of the 3DV motion stream (ablation) | PointNet++ multi-stream model · NTU RGB+D 120 and NTU RGB+D 60 | accuracy (%): 76.9 vs. 72.1; 92.5 vs. 79.4; 84.5 vs. 80.1; 95.4 vs. 85.1 | Hurt |
| Use action proposal to remove background (ablation) | 3DV-PointNet++ · N-UCLA | accuracy (%): 92.9 -> 95.3 | Helped |
| Use PointNet++ rather than C3D for 3DV (ablation) | 3DV · N-UCLA and NTU RGB+D 60 cross-view | accuracy (%): 64.5 -> 71.3; 85.0 -> 90.0 | Helped |
| Increase sampled 3DV points from 512 to 1024 to 2048 (ablation) | 3DV-PointNet++ · NTU RGB+D 120 | accuracy (%): 74.9 -> 75.7 -> 76.9; 89.0 -> 90.9 -> 92.5 | Helped |
| Vary 3DV voxel size from 25 to 75 mm (ablation) | 3DV-PointNet++ · NTU RGB+D 120 | accuracy (%): 75.9; 76.9; 76.0; 74.1; 92.0; 92.5; 91.6; 90.4 | Mixed |
| Use approximated rather than original temporal rank pooling (ablation) | 3DV · N-UCLA | accuracy: 95.8 -> 95.3 | Hurt |
| Use approximated rather than original temporal rank pooling (ablation) | 3DV · N-UCLA | time per sample: 1.12s -> 0.10s | Helped |
| Use 3DV-PointNet++ rather than the reported depth-based baselines | 3DV-PointNet++ · NTU RGB+D 120 | accuracy (%): 67.7 -> 82.4; 66.9 -> 93.5 | Helped |
| Use 3DV-PointNet++ rather than the strongest reported baseline | 3DV-PointNet++ · NTU RGB+D 60 cross-subject | accuracy (%): 89.9 vs. 88.8 | Hurt |
Closest papers by meaning
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.89)
- PA3D: Pose-Action 3D Machine for Video Recognition (CVPR 2019 · similarity 0.89)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.87)
- Scene Flow to Action Map: A New Representation for RGB-D Based Action Recognition With Convolutional Neural Networks (CVPR 2017 · similarity 0.87)
- 3DInAction: Understanding Human Actions in 3D Point Clouds (CVPR 2024 · similarity 0.85)
- Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos (CVPR 2025 · 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.