Research map / Detection and segmentation / Open-set and few-shot detection
Visual Relationship Detection Using Part-and-Sum Transformers With Composite Queries
Qi Dong, Zhuowen Tu, Haofu Liao, Yuting Zhang, Vijay Mahadevan, Stefano Soatto
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 tensor-based query instead of vector-based query (ablation) | DETR-style vanilla decoder · VRD Relationship Detection k=70 | recall@50 / recall@100: 19.97 -> 21.89; 23.08 -> 25.89 | Improved |
| replace vanilla decoder by part decoder with tensor query (ablation) | tensor query · VRD Relationship Detection k=70 | recall@50 / recall@100: 21.89 -> 25.46; 25.89 -> 29.65 | Improved |
| mix part and sum queries into composite query with vanilla decoder (ablation) | vanilla decoder · VRD Relationship Detection k=70 | recall@50 / recall@100: 19.97 -> 20.17; 23.08 -> 24.53 | Mixed |
| decode part and sum queries separately with part-sum interaction (PST decoder) (ablation) | composite query with vanilla decoder · VRD Relationship Detection k=70 | recall@50 / recall@100: 20.17 -> 26.48; 24.53 -> 31.83 | Improved |
| add factorized self-attention layer (ablation) | PST part stream · VRD Relationship Detection k=1 | recall@50 / recall@100: 22.14 -> 23.57; 26.48 -> 27.63 | Improved |
| add factorized self-attention layer (ablation) | PST part stream · VRD Relationship Detection k=70 | recall@50 / recall@100: 25.30 -> 26.48; 29.17 -> 31.83 | Improved |
| use Part-Sum summation interaction instead of self-attention interaction (ablation) | PST · VRD Relationship Detection k=70 | recall@50 / recall@100: 23.89 -> 26.48; 28.87 -> 31.83 | Improved |
| use part-only inference instead of sum-only inference (ablation) | PST · VRD Relationship Detection k=1 | recall@50 / recall@100: 22.06 -> 23.57; 25.43 -> 27.63 | Improved |
| combine part and sum predictions at inference instead of part only (ablation) | PST · VRD Relationship Detection k=1 | recall@50 / recall@100: 23.57 -> 24.34; 27.63 -> 27.01 | Mixed |
| use independent-stream decoding for part and sum queries instead of shared-stream (ablation) | PST · VRD Relationship Detection k=1 | recall@50 / recall@100: 20.04 -> 23.57; 23.10 -> 27.63 | Improved |
| increase number of vector queries from 500 to 1500 (ablation) | vector query decoder · VRD Relationship Detection k=1 | recall@50 / recall@100: 17.88 -> 18.63; 19.41 -> 21.14 | No clear effect |
| increase vector query dimension to 256x3 (ablation) | vector query decoder · VRD Relationship Detection k=1 | recall@50 / recall@100: 17.88 -> 18.65; 19.41 -> 20.94 | No clear effect |
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.
- Zoom-Net — Zoom-Net: Mining Deep Feature Interactions for Visual Relationship Recognition (ECCV 2018) · 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
- Efficient Two-Stage Detection of Human-Object Interactions With a Novel Unary-Pairwise Transformer (CVPR 2022 · similarity 0.88)
- Visual Relationship Detection with Low Rank Non-Negative Tensor Decomposition (AAAI 2020 · similarity 0.87)
- Groupwise Query Specialization and Quality-Aware Multi-Assignment for Transformer-based Visual Relationship Detection (CVPR 2024 · similarity 0.87)
- Unified Visual Relationship Detection with Vision and Language Models (ICCV 2023 · similarity 0.87)
- Scene-Graph ViT: End-to-End Open-Vocabulary Visual Relationship Detection (ECCV 2024 · similarity 0.87)
- What To Look at and Where: Semantic and Spatial Refined Transformer for Detecting Human-Object Interactions (CVPR 2022 · 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.
