Research map / 3D vision / Human pose estimation
Unified Keypoint-Based Action Recognition Framework via Structured Keypoint Pooling
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 |
|---|---|---|---|
| add object keypoints (category + contour points) to human skeletons (ablation) | Structured Keypoint Pooling with PPNv2 keypoints · Kinetics-400 | Accuracy (%): 41.2 to 49.2 (+8.0); 43.1 to 52.3 in Table 1 (+9.2) | Helped |
| use eight object contour points instead of bounding box points (ablation) | object keypoint input · Kinetics-400 | Accuracy (%): 48.6 to 49.2 (+0.6) | Helped |
| add instance-level GPB (aggregate keypoints into instance features) (ablation) | Structured Keypoint Pooling · Kinetics-400 (HRNet) | Accuracy (%) and runtime: 47.3 to 48.6 accuracy; runtime 89.5 to 7.2 ms | Helped |
| add frame-level GPB on top of instance-level GPB (ablation) | Structured Keypoint Pooling · Kinetics-400 (HRNet) | Accuracy (%) and runtime: accuracy 48.6 to 48.5 (-0.1); runtime 7.2 to 4.9 ms | Mixed |
| add MLP blocks with ConcatPool (local group pooling) instead of plain MLPs (ablation) | Structured Keypoint Pooling · Kinetics-400 (HRNet) | Accuracy (%): 44.5 to 48.5 with GPB | Helped |
| replace ConcatPool MLP block with MS-G3D temporal GCN module (ablation) | Structured Keypoint Pooling · Kinetics-400 (HRNet) | Accuracy (%): 48.5 to 45.7 (-2.8) | Hurt |
| add batch-mixing augmentation (CutMix-style inter-sample mixing) (ablation) | Pooling-Switching weakly supervised localization · UCF101-24 | Video AP@0.2 and AP@0.5: AP@0.2 60.4 to 61.8; AP@0.5 37.4 to 38.0 | Helped |
| use Structured Keypoint Pooling with skeleton + object keypoints (proposed) vs skeleton only under Mixamo-to-Kinetics domain shift (ablation) | Structured Keypoint Pooling · Mixamo to Kinetics-400 | Accuracy (%): 27.6 to 28.4 (+0.8) | Helped |
| proposed method vs. SoTA skeleton methods on skeleton-only input | Structured Keypoint Pooling · Kinetics-400 (HRNet) | Accuracy (%): 50.3 vs 47.7 (PoseConv3D) and 45.1 (MS-G3D) | Helped |
| proposed method on RWF-2000, Hockey-Fight, Crowd Violence, Movies-Fight vs SPIL | Structured Keypoint Pooling · violence recognition datasets | Accuracy (%): RWF-2000 89.3 to 93.4; Hockey 96.8 to 99.5; Crowd 94.5 to 94.7; Movies 98.5 to 99.0 | Helped |
| proposed skeleton-based method vs appearance-based on UCF101 and HMDB51 | Structured Keypoint Pooling · UCF101 / HMDB51 | Accuracy (%): UCF101 87.8 vs I3D 95.6; HMDB51 70.9 vs OmniSource 87.0 | Hurt |
Closest papers by meaning
- PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition (CVPR 2020 · similarity 0.87)
- Mining 3D Key-Pose-Motifs for Action Recognition (CVPR 2016 · similarity 0.86)
- Frame-Level Label Refinement for Skeleton-Based Weakly-Supervised Action Recognition (AAAI 2023 · similarity 0.86)
- A New Representation of Skeleton Sequences for 3D Action Recognition (CVPR 2017 · similarity 0.86)
- Prompt-Guided Zero-Shot Anomaly Action Recognition Using Pretrained Deep Skeleton Features (CVPR 2023 · similarity 0.86)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (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.