Research map / Detection and segmentation / Open-set and few-shot detection
Agglomerative Transformer for Human-Object Interaction Detection
Danyang Tu, Wei Sun, Guangtao Zhai, Wei Shen
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 extra instance-level cues (pose, spatial, object category) (ablation) | AGER · HICO-Det test (Full) | mAP: 33.07 -> 36.75 | Improved |
| add object category cue to appearance features (ablation) | AGER · HICO-Det test (Rare) | mAP: 23.81 -> 26.21 | Improved |
| add CLIP-based textual guidance to instance token clustering (ablation) | AGER · HICO-Det test (Full) | mAP: 19.82 -> 36.75 | Improved |
| use weighted CE-and-cosine similarity instead of CE alone (ablation) | AGER textual guidance loss · HICO-Det test (Full) | mAP: 19.82 -> 36.75 | Improved |
| use weighted similarity instead of cosine only (ablation) | AGER textual guidance loss · HICO-Det test (Full) | mAP: 29.80 -> 36.75 | Improved |
| add separate position embeddings for human and object centers (ablation) | AGER · HICO-Det test (Full) | mAP: 33.71 -> 36.75 | Improved |
| add cue-switch confidence threshold in cue aggregation (ablation) | AGER · HICO-Det test (Full) | mAP: 34.26 -> 36.75 | Improved |
| use 3 interaction patterns instead of 1 (ablation) | AGER interaction decoder · HICO-Det test (Full) | mAP: 34.81 -> 36.75 | Improved |
| increase patterns from 3 to 5 or 7 (ablation) | AGER interaction decoder · HICO-Det test (Full) | mAP: 36.75 -> 35.54 (5), 35.10 (7) | Hurt |
| use (16,64)/(4,8) clustering centers instead of larger numbers (ablation) | AGER · HICO-Det test (Full) | mAP: 34.42 -> 36.75 | Improved |
| extra cues when sample volumes are imbalanced (500 vs 5000) (ablation) | AGER interaction decoder · wheel vs ride bicycle subset | mAP: 61.42/15.37 -> 69.83/27.52 | Improved |
| replace QPIC-style pipeline with AGER instance-token pipeline | HOI detector efficiency · 640x640, GTX 1080Ti | FPS / GFLOPs: FPS 20.0 -> 27.2; GFLOPs 36.95 -> 33.81 | Improved |
Closest papers by meaning
- Exploring Structure-Aware Transformer Over Interaction Proposals for Human-Object Interaction Detection (CVPR 2022 · similarity 0.91)
- HOTR: End-to-End Human-Object Interaction Detection With Transformers (CVPR 2021 · similarity 0.91)
- Efficient Two-Stage Detection of Human-Object Interactions With a Novel Unary-Pairwise Transformer (CVPR 2022 · similarity 0.91)
- Iwin: Human-Object Interaction Detection via Transformer with Irregular Windows (ECCV 2022 · similarity 0.91)
- End-to-End Human Object Interaction Detection With HOI Transformer (CVPR 2021 · similarity 0.91)
- Human-Object Interaction Detection via Disentangled Transformer (CVPR 2022 · similarity 0.90)
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.
