Research map / Detection and segmentation / CNN features
Building Vision Transformers with Hierarchy Aware Feature Aggregation
Yongjie Chen, Hongmin Liu, Haoran Yin, Bin Fan
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 |
|---|---|---|---|
| insert HAFA into PVT-Tiny for classification | PVT-Tiny · ImageNet-1k val | top-1 accuracy (%): 75.1 -> 77.5 | Improved |
| insert HAFA into PVT-Tiny for semantic segmentation | PVT-Tiny · ADE20K val | mIoU (%): 35.7 -> 40.1 | Improved |
| insert HAFA into PVT-Tiny for detection | PVT-Tiny · COCO val2017 | box AP: 36.7 -> 39.8 | Improved |
| replace SIA with LAA in stage 1 (ablation) | PVT-v2-B1 HAFA · ImageNet-1k val | top-1 accuracy (%): 79.4 -> 79.7 | Improved |
| also replace SIA with LAA in stage 2 (ablation) | PVT-v2-B1 HAFA · ImageNet-1k val | top-1 accuracy (%): 79.7 -> 79.8 and latency 23.8 -> 16.3 ms | Improved |
| use LAA in all three stages (ablation) | PVT-v2-B1 · ImageNet-1k val | top-1 accuracy (%) | Hurt |
| cluster only Q with implicit positional encoding on K and V (ablation) | SIA in PVT-Tiny · ImageNet-1k val | top-1 accuracy (%): 77.1 -> 77.5 | Improved |
| cluster only Q instead of Q, K, V (ablation) | SIA in PVT-Tiny · ADE20K val | mIoU (%): 39.8 -> 40.1 | Improved |
| integrate HAFA into Swin-T versus DynamicViT-style token pruning | Swin-T · ImageNet-1k val | top-1 accuracy (%): 81.3 -> 81.7 | Improved |
| add HAFA to PVT-Small for classification | PVT-Small · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 80.1 | Improved |
| add HAFA to PVT latency | PVT · inference latency | latency (ms): +1.4 / +1.5 / +1.3 ms | Hurt |
Closest papers by meaning
- Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping (CVPR 2024 · similarity 0.85)
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.84)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.84)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.84)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.84)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.84)
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.
