Research map / Video / Action recognition
NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding
Jun Liu, Amir Shahroudy, Mauricio Pérez, Gang Wang, Ling‐Yu Duan, Alex Chichung Kot
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 |
|---|---|---|---|
| add 3D skeleton input instead of RGB or depth for cross-setup evaluation (ablation) | action recognition baselines · NTU RGB+D 120 cross-setup | accuracy: RGB 54.8% -> skeleton 57.9% | Improved |
| use RGB video instead of 3D skeleton for cross-subject evaluation (ablation) | action recognition baselines · NTU RGB+D 120 cross-subject | accuracy: 55.7% -> 58.5% | Improved |
| add depth video to RGB video (ablation) | RGB-only model · NTU RGB+D 120 cross-setup | accuracy: 54.8% -> 59.2% | Improved |
| add 3D skeleton to RGB video (ablation) | RGB-only model · NTU RGB+D 120 cross-setup | accuracy: 54.8% -> 63.1% | Improved |
| fuse all three modalities (RGB + depth + skeleton) rather than skeleton only (ablation) | skeleton model · NTU RGB+D 120 cross-subject | accuracy: 55.7% -> 64.0% | Improved |
| fuse all three modalities rather than RGB only (ablation) | RGB model · NTU RGB+D 120 cross-subject | accuracy: 58.5% -> 64.0% | Improved |
| use RGB or depth input under camera setup changes (cross-setup vs cross-subject) | VGG RGB model · NTU RGB+D 120 | accuracy: 58.5% -> 54.8% | Hurt |
| increase training set from 20% to full (ablation) | FSNet · NTU RGB+D 120 cross-setup | accuracy: 40.6% -> 62.4% | Improved |
| add semantic-relevance weighted pooling (APSR) instead of average pooling (ablation) | ST-LSTM feature generator · NTU RGB+D 120 one-shot | accuracy: 42.9% -> 45.3% | Improved |
| add learned attention over joints instead of average pooling (ablation) | one-shot ST-LSTM · NTU RGB+D 120 one-shot | accuracy: 42.9% -> 41.0% | Hurt |
| add fully connected layer on ST-LSTM outputs instead of average pooling (ablation) | one-shot ST-LSTM · NTU RGB+D 120 one-shot | accuracy: 42.9% -> 42.1% | Hurt |
| increase auxiliary training set from 20 to 100 classes (ablation) | APSR · NTU RGB+D 120 one-shot | accuracy: 29.1% -> 45.3% | Improved |
Closest papers by meaning
- NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis (CVPR 2016 · similarity 0.97)
- FLAG3D++: A Benchmark for 3D Fitness Activity Comprehension With Language Instruction (TPAMI 2025 · similarity 0.88)
- MMNet: A Model-Based Multimodal Network for Human Action Recognition in RGB-D Videos (TPAMI 2022 · similarity 0.88)
- Learning Action Recognition Model From Depth and Skeleton Videos (ICCV 2017 · similarity 0.88)
- Jointly Learning Heterogeneous Features for RGB-D Activity Recognition (TPAMI 2016 · similarity 0.88)
- 3DInAction: Understanding Human Actions in 3D Point Clouds (CVPR 2024 · 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.
