Research map / Label-efficient and robust learning / Recognition and long-tail
Learning of Visual Relations: The Devil Is in the Tails
Alakh Desai, Tz-Ying Wu, Subarna Tripathi, Nuno Vasconcelos
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 two-stage decoupled training instead of single-stage (ablation) | DT2 · VG150 SGCls | mR@100: 12.4 -> 26.7 | Improved |
| use class-balanced sampling instead of random sampling in single-stage training (ablation) | DT2 single stage · VG150 SGCls | mR@100: 11.2 -> 12.4 | Improved |
| use separate entity-CBS and predicate-CBS instead of predicate-CBS for both (ablation) | DT2 · VG150 SGCls | mR@100: 14.3 -> 26.7 | Improved |
| add ACBS distillation between P-step and E-step to independent CBS (ablation) | DT2 · VG150 SGCls | mR@100: 26.7 -> 27.5 | Improved |
| add the appearance branch (ablation) | DT2-ACBS · VG150 PredCls | mR@100: 26.8 -> 39.7 | Improved |
| add the appearance branch (ablation) | DT2-ACBS · VG150 SGCls | mR@100: 16.3 -> 27.5 | Improved |
| add the appearance branch (ablation) | DT2-ACBS · VG150 SGCls | subject-object entity accuracy: 25.77 -> 26.26 | No clear effect |
| use P-step as teacher instead of E-step (ablation) | DT2-ACBS · VG150 SGCls | mR@100: 22.0 -> 27.5 | Improved |
| replace MOTIFS with DT2-ACBS using long-tail-aware training and simple architecture | DT2-ACBS vs TDE-MOTIFS-SUM · VG150 PredCls tail classes | mR@100: 6.1 -> 38.6 | Improved |
| DT2-ACBS versus TDE-MOTIFS-SUM on head classes | DT2-ACBS · VG150 PredCls head classes | mR@100: 44.9 -> 35.1 | Hurt |
| DT2-ACBS versus PCPL on PredCls | DT2-ACBS · VG150 PredCls | mR@100: 37.8 -> 39.7 | Improved |
| DT2-ACBS versus PCPL on SGDet | DT2-ACBS · VG150 SGDet | mR@100: 11.7 -> 24.4 | Improved |
Closest papers by meaning
- RelTransformer: A Transformer-Based Long-Tail Visual Relationship Recognition (CVPR 2022 · similarity 0.90)
- Visual Distant Supervision for Scene Graph Generation (ICCV 2021 · similarity 0.90)
- Leveraging Predicate and Triplet Learning for Scene Graph Generation (CVPR 2024 · similarity 0.89)
- One-Shot Learning for Long-Tail Visual Relation Detection (AAAI 2020 · similarity 0.89)
- Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (CVPR 2024 · similarity 0.88)
- Synthetic Visual Genome (CVPR 2025 · similarity 0.87)
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.
