Research map / Architectures and efficiency / Attention and transformers
Inter2Former: Dynamic Hybrid Attention for Efficient High-Precision Interactive Segmentation
You Huang, Lichao Chen, Jiayi Ji, Liujuan Cao, Shengchuan Zhang, Rongrong Ji
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 Dynamic Hybrid Attention instead of all BSQ attention (ablation) | Inter2Former · HQSeg44K 2048 | NoC90: 5.64 -> 4.24 | Improved |
| use Dynamic Hybrid Attention instead of all BSQ attention (ablation) | Inter2Former · HQSeg44K 2048 | 5-mIoU: 90.12 -> 92.68 | Improved |
| use DHA instead of all Full Attention (ablation) | Inter2Former · HQSeg44K 1024 | NoC90: 5.38 -> 5.46 | No clear effect |
| use DHA instead of all Full Attention (ablation) | Inter2Former · HQSeg44K 2048 | NoC90: 4.24 -> 4.24 | No clear effect |
| use BSQA instead of VQA (ablation) | Inter2Former · HQSeg44K 2048 | NoC90: 4.82 -> 4.24 | Improved |
| use BSQA instead of VQA (ablation) | Inter2Former · HQSeg44K 1024 | NoC90: 5.59 -> 5.46 | Improved |
| use Dynamic Prompt Embedding instead of full prompt embedding (ablation) | Inter2Former · HQSeg44K 1024 | NoC90: 5.44 -> 5.46 | No clear effect |
| use Dynamic Local Upsampling instead of full upsampling (ablation) | Inter2Former · DAVIS 1024 | NoC95: 11.81 -> 12.72 | Hurt |
| use Dynamic Local Upsampling instead of full upsampling (ablation) | Inter2Former · DAVIS 2048 | NoC95: 7.90 -> 7.82 | No clear effect |
| increase input resolution from 1024 to 2048 (ablation) | Inter2Former · HQSeg44K | NoC90: 5.46 -> 4.24 | Improved |
| use CPU-optimized HMoE instead of vanilla MoE (ablation) | Inter2Former · CPU latency | latency: 56-85% latency reduction | Improved |
| use SAM distillation protocol instead of COCO+LVIS training (ablation) | Inter2Former · HQSeg44K 2048 | NoC90: 4.58 -> 4.24 | Improved |
Closest papers by meaning
- FocSAM: Delving Deeply into Focused Objects in Segmenting Anything (CVPR 2024 · similarity 0.90)
- Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts (CVPR 2024 · similarity 0.88)
- SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model (CVPR 2025 · similarity 0.87)
- Order-aware Interactive Segmentation (ICLR 2025 · similarity 0.86)
- RClicks: Realistic Click Simulation for Benchmarking Interactive Segmentation (NeurIPS 2024 · similarity 0.86)
- AGILE3D: Attention Guided Interactive Multi-object 3D Segmentation (ICLR 2024 · similarity 0.86)
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.
