Research map / Detection and segmentation / Open-set and few-shot detection
HOIMamba: Efficient Mamba-based Disentangled Progressive Learning for HOI Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HOIMamba, a Mamba-based HOI detection decoder with cascaded LoRA branches and cross-enhance/context-propagation blocks that progressively aggregate interaction features, reaching SOTA on HICO-DET and V-COCO with fewer parameters.
Problem
Human-object interaction detection struggles to trade off efficiency, multi-task decoupling, and collaborative learning across single-, two-, and three-branch Transformer decoders, and fails on rare and complex interaction categories.
Approach
Builds a decoder from cascaded LoRA branches for human, object and interaction tasks with per-task routers, and adds Mamba-based Cross-enhance Mamba (CEM) and Detection Context Propagation (DCP) blocks that progressively aggregate four levels of interaction features and exchange cross-task context.
Key results
- HOIMamba reaches 41.51 full mAP on HICO-DET with ResNet-50, 41.77 with ResNet-101 and 47.66 with Swin-Large backbones, claimed as SOTA.
- Ablation on HICO-DET (ResNet-50) shows progressive learning (PL) adds +8.25 rare mAP and DCP adds +1.10 full mAP.
- Efficiency: 40.31M params, 31.15 GFLOPs, 27.26 FPS versus QPIC's 42.35M, 36.91 GFLOPs, 20.04 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace decoder self-attention and cross-attention with CEM blocks (ablation) | QPIC-based HOIMamba · HICO-DET Default | mAP (Rare): +1.54 | Helped |
| Add three-branch task disentangling (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Rare): +4.84 | Helped |
| Add human and object branch features as interaction priors (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Rare): +1.35 | Helped |
| Add human-object interaction in the second CEM block (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Rare) | Helped |
| Add task-specific routers (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Full): +1.60 | Helped |
| Add comprehensive progressive learning (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Rare): +8.25 | Helped |
| Add Detection Context Propagation (ablation) | HOIMamba ablation model · HICO-DET Default | mAP (Full): +1.10 | Helped |
| Set CEM φ1=0.50 and φ2=0.75 rather than the other tested settings (ablation) | HOIMamba · HICO-DET Default | mAP (Full) | Helped |
| Apply DCP to both human and object branches instead of neither (ablation) | HOIMamba · HICO-DET Default | mAP (Full): 40.41 -> 41.51 | Helped |
Limitations
- Fails on some cases due to limited visual representation (e.g., missing <human, pet, cat>).
- Rare-category gains rely on the progressive learning design, and the paper reports limited analysis beyond two benchmarks.
Experimental setup
Datasets: HICO-DET V-COCO MS-COCO
Benchmarks: HICO-DET V-COCO
Metrics: mAP FPS GFLOPs parameter count
Baselines: QPIC UPT MUREN HybHOI DPAD CDN GEN-VLKT RmLR SCTC FGAHOI PViC MP-HOI
Builds on: QPIC MUREN Mamba Vision Mamba LoRA DETR
Closest papers by meaning
- Exploring Pose-Aware Human-Object Interaction via Hybrid Learning (CVPR 2024 · similarity 0.89)
- HORP: Human-Object Relation Priors Guided HOI Detection (CVPR 2025 · similarity 0.89)
- Toward Open-Set Human Object Interaction Detection (AAAI 2024 · 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 0.88)
- Learning Human-Object Interaction as Groups (NeurIPS 2025 · 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.