Research map / Video / Video segmentation and tracking
Beyond Euclidean: Dual-Space Representation Learning for Weakly Supervised Video Violence 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 DSRL, a dual-space method combining Euclidean and Lorentz hyperbolic representations with a hyperbolic-Dirichlet-energy-constrained GCN and cross-space attention for weakly supervised video violence detection on XD-Violence and UCF-Crime.
Problem
Weakly supervised video violence detection struggles to separate ambiguous violence (normal events visually similar to violent ones) when representations are learned only in Euclidean space or only in hyperbolic space.
Approach
Jointly learns Euclidean and Lorentz-hyperbolic features; HE-GCN replaces HGCN's fixed-threshold node selection with layer-sensitive hyperbolic association degrees (dynamic thresholds driven by hyperbolic Dirichlet energy), and DSI uses Lorentzian-similarity cross-space attention to exchange information between the two spaces.
Key results
- On XD-Violence multimodal setting, DSRL reaches 87.61% AP, versus 85.67% for HyperVD (hyperbolic-only) and 83.40% for MACIL-SD.
- On XD-Violence unimodal setting, DSRL reaches 82.01% AP, the best reported in the table (MSL 78.28%, UMIL 81.66%).
- On UCF-Crime, DSRL reaches 86.38% AUC, comparable to MGFN (86.98%) and UMIL (86.75%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine Euclidean and hyperbolic representations with DSRL | Weakly supervised VVD · XD-Violence (multimodal) | AP: 85.67 -> 87.61 | Helped |
| combine Euclidean and hyperbolic representations with DSRL | Weakly supervised VVD · XD-Violence (unimodal) | AP: 81.66 -> 82.01 | Helped |
| combine Euclidean and hyperbolic representations with DSRL | Weakly supervised VVD · UCF-Crime | AUC: 86.98 -> 86.38 | Hurt |
| add HE-GCN instead of HGCN's fixed-threshold node selection (ablation) | DSRL with GCN and hyperbolic representation · XD-Violence (multimodal) | AP: 85.01 -> 86.46 | Helped |
| add HE-GCN instead of HGCN's fixed-threshold node selection (ablation) | DSRL with GCN and hyperbolic representation · XD-Violence (unimodal) | AP: 77.93 -> 79.70 | Helped |
| add HE-GCN to Euclidean GCN and concatenate the representations (ablation) | DSRL · XD-Violence (multimodal) | AP: 84.04 -> 86.46 | Helped |
| add HE-GCN to Euclidean GCN and concatenate the representations (ablation) | DSRL · XD-Violence (unimodal) | AP: 77.95 -> 79.70 | Helped |
| add DSI with Lorentzian cross-space attention instead of simple concatenation (ablation) | DSRL with HE-GCN · XD-Violence (multimodal) | AP: 86.46 -> 87.61 | Helped |
| add DSI with Lorentzian cross-space attention instead of simple concatenation (ablation) | DSRL with HE-GCN · XD-Violence (unimodal) | AP: 79.70 -> 82.01 | Helped |
| use Lorentzian rather than cosine similarity in DSI (ablation) | DSRL with DSI · XD-Violence (multimodal) | AP: 86.91 -> 87.61 | Helped |
| vary DSI lambda from 0.1 to 0.9 (ablation) | DSRL with DSI · XD-Violence (multimodal) | AV-AP: 87.14, 87.38, 87.1, 87.32, 86.53, 87.2, 87.4, 87.61, 87.23 | Mixed |
| vary DSI alpha from 0.1 to 0.9 (ablation) | DSRL with DSI · XD-Violence (multimodal) | AV-AP: 87.18, 87.16, 87.61, 87.21, 87.26, 86.91, 87.42, 86.59, 87.51 | Mixed |
Limitations
- Uses only basic audio information, possibly missing detailed audio semantics.
- Gains on UCF-Crime are only comparable to, not better than, the best Euclidean methods.
- Several hyperparameters (beta, gamma, alpha, lambda) are hand-set and sensitivity to them is only partly explored.
Experimental setup
Datasets: XD-Violence UCF-Crime
Benchmarks: XD-Violence UCF-Crime
Metrics: frame-level AP frame-level AUC
Baselines: Sultani et al. Wu et al. RTFM MSL MGFN UMIL CU-Net HL-Net Pang et al. Zhang et al. MACIL-SD HyperVD
Builds on: HyperVD HGCN multiple-instance learning Lorentz model
Closest papers by meaning
- Bullying10K: A Large-Scale Neuromorphic Dataset towards Privacy-Preserving Bullying Recognition (NeurIPS 2023 · similarity 0.84)
- Dual Memory Units with Uncertainty Regulation for Weakly Supervised Video Anomaly Detection (AAAI 2023 · similarity 0.84)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · similarity 0.82)
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · similarity 0.81)
- Hyperbolic Category Discovery (CVPR 2025 · similarity 0.80)
- Towards Effective, Efficient and Unsupervised Social Event Detection in the Hyperbolic Space (AAAI 2025 · similarity 0.80)
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.