Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Social Relation Representation for Human Group Detection
Jiacheng Li, Ruize Han, Haomin Yan, Zekun Qian, Wei Feng, Song Wang
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 self-supervised behavior-simulator first-stage training (ablation) | two-stage group detector · PANDA group detection | F1 (%): 44.6 -> 53.2 | Improved |
| add self-supervised behavior-simulator first-stage training (ablation) | two-stage group detector · JRDB-Group group detection | F1 (%): 48.8 -> 56.9 | Improved |
| add fine-tuning of the embedding network in stage 2 (ablation) | two-stage group detector · PANDA group detection | F1 (%): 51.1 -> 53.2 | Improved |
| add fine-tuning of the embedding network in stage 2 (ablation) | two-stage group detector · JRDB-Group group detection | F1 (%): 49.2 -> 56.9 | Improved |
| add self-supervised first stage when only 10% of labels are used (ablation) | two-stage group detector · PANDA group detection (10% train) | F1 (%): 27.6 -> 40.9 | Improved |
| add self-supervised first stage; 10% labels vs 50% labels without it (ablation) | two-stage group detector · PANDA group detection (10% vs 50% train) | F1 (%): 39.8 -> 40.9 | Improved |
| freeze the embedding network (no fine-tuning) with only 10% labels (ablation) | two-stage group detector · JRDB-Group group detection (10% train) | F1 (%): 37 and 43.8 | Improved |
| use proposed method instead of best baseline ARG | ARG · PANDA group detection | F1 (%): 25.4 -> 53.2 | Improved |
| use proposed method instead of JRDB-Group baseline | JRDB-Group baseline · JRDB-Group group detection | F1 (%): 38.4 -> 56.9 | Improved |
| use proposed method instead of simple distance baselines under IOU^GM | Dis.Mat + label propagation · JRDB-Group group detection | IOU^GM (%): 12.9 -> 32.5 | Improved |
Closest papers by meaning
- Joint Learning of Social Groups, Individuals Action and Sub-group Activities in Videos (ECCV 2020 · similarity 0.89)
- Social Scene Understanding: End-To-End Multi-Person Action Localization and Collective Activity Recognition (CVPR 2017 · similarity 0.87)
- JRDB-Act: A Large-Scale Dataset for Spatio-Temporal Action, Social Group and Activity Detection (CVPR 2022 · similarity 0.86)
- Learning Human-Object Interaction as Groups (NeurIPS 2025 · similarity 0.85)
- Part-Aware Bottom-Up Group Reasoning for Fine-Grained Social Interaction Detection (NeurIPS 2025 · similarity 0.85)
- Behavioral Geometric Supervision Aligns Video Foundation Models with Human Social Perception (NeurIPS 2026 · similarity 0.84)
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.
