Research map / Vision-language and multimodal / CLIP and zero-shot
ART: Adaptive Relation Tuning for Generalized Relation Prediction
Gopika Sudhakaran, Hikaru Shindo, Patrick Schramowski, Simone Schaub-Meyer, Kristian Kersting, Stefan Roth
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 |
|---|---|---|---|
| tune on adaptively selected 12% of data (ART) instead of balanced random sampling | InstructBLIP Vicuna-7B · GQA | mR@20: 13.2 -> 18.9 | Improved |
| use ART instead of random sampling on Recall | InstructBLIP Vicuna-7B · GQA | R@20: 59.0 -> 40.1 | Hurt |
| tune on random 12% subset instead of full VG train data (ablation) | InstructBLIP Vicuna-7B · VG | mR@20: 14.1 -> 19.6 | Improved |
| use balanced random sampling instead of random (ablation) | InstructBLIP Vicuna-7B · VG | mR@20: 19.6 -> 42.4 | Improved |
| add adaptive sampling (hFN, tFP, tFN, hTP) on top of balanced sampling (ablation) | ART · VG | mR@20: 42.4 -> 46.4 | Improved |
| select by TP/FN/FP-specific criteria instead of entropy-only (ablation) | ART · VG | mR@20: 44.6 -> 46.4 | Improved |
| add high-level relation types to the prompt (ablation) | ART instruction set · VG | mR@20: 37.5 -> 39.4 | Improved |
| replace random negatives with one counter-negative (ablation) | ART instruction set · VG | mR@20: 40.9 -> 46.4 | Improved |
| increase counter-negatives per sample from 1 to 2 (ablation) | ART instruction set · VG | mR@20: 46.4 -> 40.3 | Hurt |
| use adaptive thresholds instead of fixed thresholds (ablation) | ART · VG | mR@20: 44.8 -> 46.4 (best fixed vs adaptive) | Improved |
| use ART-predicted scene graphs in DeiSAM instead of ground-truth graphs | DeiSAM · Deictic segmentation VG (w/o bbox) | mAP: 4.77 -> 25.07 | Improved |
| ART on detected RAM boxes vs random sampling | InstructBLIP Vicuna-7B · OI-v4 with RAM boxes | mR@20: 6.7 -> 24.0 | Improved |
Closest papers by meaning
- Zero-shot Visual Relation Detection via Composite Visual Cues from Large Language Models (NeurIPS 2023 · similarity 0.91)
- Visual Relationship Detection With Internal and External Linguistic Knowledge Distillation (ICCV 2017 · similarity 0.91)
- Localize, Assemble, and Predicate: Contextual Object Proposal Embedding for Visual Relation Detection (AAAI 2020 · similarity 0.91)
- The All-Seeing Project V2: Towards General Relation Comprehension of the Open World (ECCV 2024 · similarity 0.90)
- Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (CVPR 2024 · similarity 0.89)
- Relation-aware Hierarchical Prompt for Open-vocabulary Scene Graph Generation (AAAI 2025 · similarity 0.89)
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.
