Research map / Detection and segmentation / Open-set and few-shot detection
Unified Visual Relationship Detection with Vision and Language Models
Long Zhao, Liangzhe Yuan, Boqing Gong, Yin Cui, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, Ting Liu
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 |
|---|---|---|---|
| use cascade two-stage training instead of one-stage end-to-end (ablation) | UniVRD · HICO-DET | mAP: -5.05 mAP when using one-stage | Improved |
| add Objects365, COCO and VG object detection data (ablation) | UniVRD · HICO-DET | mAP: -3.96 / -2.67 / -3.16 mAP when removed | Improved |
| add mosaics augmentation (ablation) | UniVRD · HICO-DET | mAP: -1.76 when removed | Improved |
| use per-class PNMS instead of vanilla or none (ablation) | UniVRD · HICO-DET | mAP: -2.01 no PNMS, -1.24 vanilla PNMS | Improved |
| freeze object detector and text encoder in the second stage (ablation) | UniVRD · HICO-DET | mAP: -1.03 (fine-tune detector), -0.92 (fine-tune text encoder) | Improved |
| use CLIP prompt templates or prompt ensemble for relationship text queries (ablation) | UniVRD · HICO-DET | mAP: 0.00 / +0.01 | No clear effect |
| train a unified model on HICO-DET, V-COCO and VG instead of dataset-specific | UniVRD · V-COCO | mAP Scenario 1: +5.21 (ViT-B/32) to +5.39 (ViT-L/14) | Improved |
| train unified model instead of dataset-specific on Visual Genome | UniVRD · Visual Genome | mR@50: -0.01 to -0.32 mR@50 | Hurt |
| scale backbone from ViT-B/32 to ViT-H/14 (LiT) | UniVRD · HICO-DET | mAP: 29.38 -> 38.07 | Improved |
| adopt UniVRD bottom-up VLM detector vs best prior bottom-up HOI detector | UniVRD · HICO-DET | mAP: +14.26 mAP | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HOTR — HOTR: End-to-End Human-Object Interaction Detection With Transformers (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection (ECCV 2024 · similarity 0.91)
- Large-Scale Visual Relationship Understanding (AAAI 2019 · similarity 0.90)
- HCVRD: A Benchmark for Large-Scale Human-Centered Visual Relationship Detection (AAAI 2018 · similarity 0.90)
- Visual Relationship Detection with Language Priors (ECCV 2016 · similarity 0.90)
- Towards Flexible Visual Relationship Segmentation (NeurIPS 2024 · similarity 0.89)
- Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (CVPR 2024 · similarity 0.88)
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.
