Research map / Label-efficient and robust learning / Recognition and long-tail
Hyperpixel Flow: Semantic Correspondence With Multi-Layer Neural Features
Juhong Min, Jongmin Lee, Jean Ponce, Minsu Cho
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 regularized Hough matching instead of nearest neighbor (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.1: 69.0 -> 81.4 | Improved |
| use RHM with similarity exponent d=2 instead of d=1 (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.1: 81.4 -> 84.4 | Improved |
| raise exponent d from 3 to 5 (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.1: 84.8 -> 84.5 | No clear effect |
| use beam-searched layers instead of all layers (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.1: 74.5 -> 84.8 | Improved |
| use beam-searched layers instead of all layers (ablation) | HPF ResNet-50 · PF-PASCAL | PCK@0.1: 70.1 -> 83.4 | Improved |
| use layer search instead of random 8 layers (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.05: 44.5 -> 60.1 | Improved |
| use layer search instead of random 8 layers (ablation) | HPF ResNet-101 · PF-WILLOW | PCK@0.1: 62.4 -> 74.4 | Improved |
| search layers with only 1 pair per class instead of all 308 validation pairs (ablation) | HPF ResNet-101 · PF-PASCAL | PCK@0.1: 84.8 -> 83.9 | No clear effect |
| use FCN backbone pretrained on PASCAL VOC instead of ResNet-101 | HPF · PF-PASCAL | PCK@0.1: 84.8 -> 88.3 | Improved |
| use ResNet-101 instead of ResNet-50 | HPF · PF-PASCAL | PCK@0.1: 83.4 -> 84.8 | Improved |
| use fewer layers (4,7,11,12,13) with ResNet-50 for speed | HPF · PF-PASCAL | PCK@0.1 and time: 83.4 -> 81.1 | Mixed |
| finetune NC-Net on SPair-71k | NC-Net · SPair-71k | PCK@0.1 (bbox): 26.4 -> 20.1 | Hurt |
Closest papers by meaning
- Learning to Compose Hypercolumns for Visual Correspondence (ECCV 2020 · similarity 0.92)
- Convolutional Hough Matching Networks for Robust and Efficient Visual Correspondence (TPAMI 2023 · similarity 0.90)
- Convolutional Hough Matching Networks (CVPR 2021 · similarity 0.89)
- Universal Correspondence Network (NeurIPS 2016 · similarity 0.87)
- Deep Semantic Feature Matching (CVPR 2017 · similarity 0.87)
- Pixel-level Semantic Correspondence through Layout-aware Representation Learning and Multi-scale Matching Integration (CVPR 2024 · similarity 0.87)
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.
