Research map / Video / Video segmentation and tracking
Dual Conditioned Motion Diffusion for Pose-Based Video Anomaly Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dual Conditioned Motion Diffusion (DCMD), combining reconstruction and diffusion-based prediction branches for pose-based video anomaly detection, with a United Association Discrepancy regularizer and mask completion at inference.
Problem
Pose-based video anomaly detection, where existing reconstruction-based methods miss abnormal deviations and prediction-based methods are prone to noise.
Approach
Dual conditioning of a diffusion model on a noised history motion (conditioned motion) and an autoencoder history embedding (conditioned embedding), plus a symmetric KL-based United Association Discrepancy regularizer over time and self-attention associations and a mask completion step at inference.
Key results
- DCMD reaches 78.6% AUC on HR-STC, 90.0% on HR-Avenue, 69.0% on HR-UBnormal and 69.0% on UBnormal, outperforming MoCoDAD by 1.0, 1.0, 0.6 and 0.7 points.
- Ablation on HR-Avenue: removing conditioned embedding drops AUC by 1.3 points, removing UAD by 0.9, removing mask completion by 0.9.
- Removing the DCT drops HR-Avenue AUC by 0.8 points to 89.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine reconstruction and prediction branches | DCMD · HR-STC, HR-Avenue, HR-UBnormal, UBnormal | AUC (%): 77.1, 87.4, 66.4, 66.2 -> 78.6, 90.0, 69.0, 69.0 | Helped |
| Add UAD (ablation) | DCMD · HR-Avenue | AUC (%): 89.1 -> 90.0 | Helped |
| Add conditioned embedding (CE) (ablation) | DCMD · HR-Avenue | AUC (%): 88.7 -> 90.0 | Helped |
| Add mask completion (MC) (ablation) | DCMD · HR-Avenue | AUC (%): 89.1 -> 90.0 | Helped |
| Add UAD and conditioned embedding (ablation) | DCMD · HR-Avenue | AUC (%): 87.2 -> 90.0 | Helped |
| Add UAD and mask completion (ablation) | DCMD · HR-Avenue | AUC (%): 87.8 -> 90.0 | Helped |
| Add conditioned embedding and mask completion (ablation) | DCMD · HR-Avenue | AUC (%): 87.0 -> 90.0 | Helped |
| Add UAD, conditioned embedding, and mask completion to pure-noise diffusion (ablation) | DCMD · HR-Avenue | AUC (%): 86.7 -> 90.0 | Helped |
| Use DCT (ablation) | DCMD · HR-Avenue | AUC (%): 89.2 -> 90.0 | Helped |
| Use history length 3 instead of 2, 4, or 5 (ablation) | DCMD · HR-STC, HR-Avenue | AUC: 77.0, 77.8, 77.7 -> 78.6; 86.7, 88.0, 87.1 -> 89.9 | Mixed |
| Use future length 4 instead of 3 or 5 (ablation) | DCMD · HR-STC, HR-Avenue | AUC: 77.9, 77.8 -> 78.1; 89.9, 88.5 -> 90.0 | Mixed |
| Set UAD loss weight λ to 0.01 rather than 0, 0.02, or 0.05 (ablation) | DCMD · HR-STC, HR-Avenue | AUC: 77.6, 77.8, 75.6 -> 77.9; 89.1, 88.6, 87.0 -> 90.0 | Mixed |
Limitations
- Relies on skeletons detected by pose estimation, so segments without detectable skeletons are removed from evaluation.
- Sensitive to the UAD regularization weight; larger lambda values degrade performance noticeably.
Experimental setup
Datasets: ShanghaiTech Campus (HR-STC) CUHK Avenue (HR-Avenue) HR-UBnormal UBnormal
Benchmarks: HR-STC HR-Avenue HR-UBnormal UBnormal
Metrics: AUC
Baselines: MPED-RNN GEPC PoseCVAE MoCoDAD COSKAD TrajREC
Builds on: MoCoDAD Anomaly Transformer STSAE FiLM
Closest papers by meaning
- Video Anomaly Detection with Motion and Appearance Guided Patch Diffusion Model (AAAI 2025 · similarity 0.89)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.86)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
- Action Detection via an Image Diffusion Process (CVPR 2024 · similarity 0.85)
- Noise-Resistant Video Anomaly Detection via RGB Error-Guided Multiscale Predictive Coding and Dynamic Memory (CVPR 2025 · similarity 0.85)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · 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.