Research map / Label-efficient and robust learning / OOD and anomaly detection
Graph Embedded Pose Clustering for Anomaly Detection
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 |
|---|---|---|---|
| Use patch embeddings instead of keypoint coordinates (ablation) | Ours on ShanghaiTech Campus · ShanghaiTech Campus | frame-level AUC: 0.752 -> 0.761 | Helped |
| Use spatial attention graph convolution rather than physical-adjacency GCN (ablation) | Ours with patch features on ShanghaiTech Campus · ShanghaiTech Campus | frame-level AUC: 0.749 -> 0.761 | Helped |
| Use spatial attention graph convolution rather than physical-adjacency GCN (ablation) | Ours with pose coordinates on ShanghaiTech Campus · ShanghaiTech Campus | frame-level AUC: 0.750 -> 0.753 | Helped |
| Use Dirichlet scoring instead of maximum softmax scoring (ablation) | K-means init, DEC on Kinetics-250 Music split · Kinetics-250, Few vs. Many, Music split, K=20 | AUC: 0.71 -> 0.82 | Helped |
| Add DEC fine-tuning after K-means initialization (ablation) | Kinetics-250 Music split with Dirichlet scoring · Kinetics-250, Few vs. Many, Music split, K=20 | AUC: 0.59 -> 0.82 | Helped |
| Use K-means rather than random clustering initialization (ablation) | DEC with Dirichlet scoring on Kinetics-250 Music split · Kinetics-250, Few vs. Many, Music split, K=20 | AUC: 0.52 -> 0.82 | Helped |
| Increase cluster count from 5 to 20 (ablation) | K-means init, DEC, Dirichlet scoring on Kinetics-250 Music split · Kinetics-250, Few vs. Many, Music split | AUC: 0.68 -> 0.82 | Helped |
| Use the proposed method rather than the strongest unsupervised baseline | Ours on NTU-RGB+D · NTU-RGB+D, Few vs. Many, Random | AUC: 0.60 -> 0.73 | Helped |
| Use the proposed method rather than the strongest unsupervised baseline | Ours on Kinetics-250 · Kinetics-250, Many vs. Few, Meaningful | AUC: 0.58 -> 0.74 | Helped |
| Train on meaningful rather than random action splits | Ours on NTU-RGB+D, Few vs. Many · NTU-RGB+D, Few vs. Many | AUC: 0.73 -> 0.82 | Helped |
| Train with 5% abnormal examples mixed into normal training data (ablation) | Ours on NTU-RGB+D splits · NTU-RGB+D noisy training data | AUC: less than 10% performance loss | Hurt |
| Train with 20% abnormal examples mixed into normal training data (ablation) | Ours on NTU-RGB+D splits · NTU-RGB+D noisy training data | AUC: 20% | Hurt |
Closest papers by meaning
- Graph Convolutional Label Noise Cleaner: Train a Plug-And-Play Action Classifier for Anomaly Detection (CVPR 2019 · similarity 0.86)
- UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection (CVPR 2022 · similarity 0.85)
- PAD: A Dataset and Benchmark for Pose-agnostic Anomaly Detection (NeurIPS 2023 · similarity 0.85)
- Dual Conditioned Motion Diffusion for Pose-Based Video Anomaly Detection (AAAI 2025 · similarity 0.84)
- Video Anomaly Detection with Motion and Appearance Guided Patch Diffusion Model (AAAI 2025 · similarity 0.83)
- Error Detection in Egocentric Procedural Task Videos (CVPR 2024 · similarity 0.83)
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.