Research map / Image generation / Image restoration
Beyond Predictive Resampling: Learning Input-Agnostic Downsampling for Efficient Aligned Vision Recognition
Kai Zhao, Liting Ruan, Haoran Jiang, Xiaoqiang Zhu, Xianchao Zhang, Dan Zeng
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 |
|---|---|---|---|
| replace uniform downsampling with learned input-agnostic flow downsampling | ArcFace · LFW at 56x56 | accuracy (%): 85.16 -> 90.59 | Improved |
| replace uniform downsampling with learned input-agnostic flow downsampling | ArcFace · IJB-B at 56x56, TAR@FAR 1e-4 | TAR (%): 81.25 -> 86.72 | Improved |
| use learned shared flow field instead of predicting per-image importance (Recasens, Jin) (ablation) | face and palmprint recognition at 56x56 · IJB-B and CASIA | TAR (%): IJB-B TAR@1e-4 (ours) 86.17 -> 86.72 | Improved |
| use learned shared grid instead of predicted grid for Recasens and Jin importance-map methods (ablation) | Recasens, Jin · CASIA palmprint 56x56 | TAR (%): Recasens 82.54 (Pred) vs 82.37 (Learn) | Mixed |
| use flow-based representation instead of importance maps (ablation) | learned and predicted downsampling · IJB-B 56x56 | TAR (%): Learn: Jin 86.01 -> Ours 86.72 at FAR 1e-4 | Improved |
| remove per-input importance prediction (input-agnostic flow) | adaptive downsampling · throughput at 56x56 | FPS: 8,539 -> 11,849 | Improved |
| use learned flow downsampling instead of uniform | ResNet-50 + ArcFace · CASIA palmprint 56x56, TAR@1e-4 | TAR (%): 83.39 -> 87.37 | Improved |
| downsample to 96x96 with learned flow instead of full 112x112 input | ArcFace · IJB-B TAR@FAR 1e-4 | TAR (%): 94.09 -> 93.14 | Mixed |
| learned flow downsampling versus uniform on palmprint sets with low absolute accuracy | ResNet-50 + ArcFace · MPD, TAR@FAR 1e-4 at 112 vs ours at 56 | TAR (%): Ours 52.90 vs Recasens 52.65 at 56x56 | No clear effect |
Closest papers by meaning
- Learning to Downsample for Segmentation of Ultra-High Resolution Images (ICLR 2022 · similarity 0.85)
- Jointly Aligning Millions of Images With Deep Penalised Reconstruction Congealing (ICCV 2019 · similarity 0.84)
- Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling (CVPR 2026 · similarity 0.84)
- Learning to Zoom: a Saliency-Based Sampling Layer for Neural Networks (ECCV 2018 · similarity 0.84)
- Kernel Aware Resampler (CVPR 2023 · similarity 0.84)
- Fix Your Downsampling ASAP! Aliasing and Sinc Artifact Free Pooling in the Fourier Domain (IJCV 2026 · 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.
