Research map / Vision-language and multimodal / CLIP and zero-shot
CrossHOI-Bench: A Unified Benchmark for HOI Evaluation across Vision-Language Models and HOI-Specific Methods
Qinqian Lei, Bo Wang, Robby T. Tan
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 |
|---|---|---|---|
| evaluate with the target person's box given instead of requiring detection (Setting 2 vs 1) | Qwen2.5-VL-7B · CrossHOI-Bench | Instance-F1: 30.53 -> 57.25 | Improved |
| evaluate with the target person's box given instead of requiring detection (Setting 2 vs 1) | InternVL3-8B · CrossHOI-Bench | Instance-F1: 4.96 -> 52.35 | Improved |
| evaluate with the target person's box given instead of requiring detection (Setting 2 vs 1) | Qwen3-VL-30B-Instruct · CrossHOI-Bench | Instance-F1: 0.25 -> 64.00 | Improved |
| evaluate with the target person's box given instead of requiring detection (Setting 2 vs 1) | Qwen2.5-VL-32B · CrossHOI-Bench | Instance-F1: 52.94 -> 69.52 | Improved |
| require recognizing interactions of all people (Setting 3 vs 2) instead of the target person only | VLMs · CrossHOI-Bench | EM: decreases by 5-10 | Hurt |
| require recognizing interactions of all people (Setting 3 vs 2) instead of the target person only | InternVL3-8B · CrossHOI-Bench | average precision: decreases by 9.1% | Mixed |
| remove overly simple head-class scenes from HICO-DET test set | CrossHOI-Bench test set · HICO-DET train vs test | KL divergence: 0.088 -> 0.629 | Improved |
| use top-5 selection for HOI-specific models rather than threshold filtering (ablation) | HOI-specific methods · CrossHOI-Bench | matching strategy | Improved |
| compare with HOI-specific methods agreement on coarse-screened negatives | HOI-specific methods vs VLMs · CrossHOI-Bench negatives | agreement: ~99% vs. 95-97% | Improved |
| use the CrossHOI-Bench subset rather than the full HICO-DET test set | VLMs and HOI-specific methods · HICO-DET-based evaluation | performance | Hurt |
| fine-tune Qwen2.5-VL-7B on HICO-DET-based train set (SFT) | Qwen2.5-VL-7B · V-COCO and SWiG-HOI sub-benchmarks | generalization | Improved |
Closest papers by meaning
- HOLa: Zero-Shot HOI Detection with Low-Rank Decomposed VLM Feature Adaptation (ICCV 2025 · similarity 0.90)
- Bilateral Collaboration with Large Vision-Language Models for Open Vocabulary Human-Object Interaction Detection (ICCV 2025 · similarity 0.90)
- EZ-HOI: VLM Adaptation via Guided Prompt Learning for Zero-Shot HOI Detection (NeurIPS 2024 · similarity 0.90)
- LINK: Learning Instance-level Knowledge from Vision-Language Models for Human-Object Interaction Detection (ICLR 2026 · similarity 0.89)
- Open-World Human-Object Interaction Detection via Multi-modal Prompts (CVPR 2024 · similarity 0.89)
- Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection (CVPR 2024 · 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.
