Research map / Detection and segmentation / CNN features
BiHashFormer: Hash-Driven Dual-Branch Transformer for Efficient Object Detection in HRW Shots
Jingchen Huang, Wenxi Li, Chenyang Lyu, Moran Liu, Haozhe Lin, Yuchen Guo
Read the paperCheck your idea against 137,217 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 BiHashFormer, a hash-driven dual-branch transformer for efficient object detection in high-resolution wide (HRW) shots with sparse foreground.
Problem
Object detection in gigapixel-scale High-Resolution Wide (HRW) shots, where foreground is extremely sparse and interleaved with background, causing traditional detectors and attention to be inefficient and inaccurate.
Approach
A hash-driven dual-branch transformer that uses hashing to route/select foreground-relevant regions so attention avoids background, improving both efficiency and accuracy (details beyond the abstract not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SparseFormer backbone with BiHashFormer | DINO head · PANDA | AP50 / GFLOPs: AP50 0.780 -> 0.810; GFLOPs 75.71 -> 47.31 | Improved |
| replace Swin-T backbone with BiHashFormer | DINO head · PANDA | AP50 / GFLOPs: AP50 0.606 -> 0.810; GFLOPs 132.84 -> 47.31 | Improved |
| disable the compensation branch at inference (ablation) | BiHashFormer · PANDA | AP small: 0.591 -> 0.287 | Hurt |
| replace HashMiner with pooled classical attention for compensation (ablation) | BiHashFormer · PANDA | small-object efficiency | Hurt |
| use random, MLP or gradient-based selection instead of HashMiner (ablation) | compensation branch · PANDA | AP small | Hurt |
| swap order to K selects 50% Q instead of Q selects 50% K (ablation) | HashAttention · PANDA | AP50 | Hurt |
| use HashAttention instead of classical attention (Q to 100% K) (ablation) | BiHashFormer · PANDA | AP50 and cost | Improved |
| increase Hamming dimension k from 16 to 32 (ablation) | BiHashFormer · PANDA | AP50: 0.801 -> 0.810 | Improved |
| increase k from 32 to 64 (ablation) | BiHashFormer · PANDA | AP50: 0.810 -> 0.805 | No clear effect |
| use 3 prototype codes instead of 1 (ablation) | HashMiner · PANDA | AP small: 0.498 -> 0.591 | Improved |
| use 5 prototype codes instead of 3 (ablation) | HashMiner · PANDA | AP small: 0.591 -> 0.582 | No clear effect |
| use dynamic foreground-proportion RoI selection instead of fixed ratio (ablation) | BiHashFormer vs SparseFormer · PANDA | AP50 | Improved |
Closest papers by meaning
- BiFormer: Vision Transformer With Bi-Level Routing Attention (CVPR 2023 · similarity 0.86)
- Efficient Decoder-Free Object Detection with Transformers (ECCV 2022 · similarity 0.85)
- Q-DETR: An Efficient Low-Bit Quantized Detection Transformer (CVPR 2023 · similarity 0.85)
- AQ-DETR: Low-Bit Quantized Detection Transformer with Auxiliary Queries (AAAI 2024 · similarity 0.84)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.84)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.83)
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.
