Research map / Detection and segmentation / CNN features
Learning to Upsample by Learning to Sample
Wenze Liu, Hao Lu, Hongtao Fu, Zhiguo Cao
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 bilinear instead of nearest initial sampling positions (ablation) | DySample · ADE20K (SegFormer-B1) | mIoU: 41.9 -> 42.1 | Improved |
| multiply offset by static scope factor 0.25 (ablation) | DySample · ADE20K (SegFormer-B1) | mIoU: 42.1 -> 42.4 | Improved |
| strictly constrain offset scope with tanh instead of soft factor | DySample · detection / segmentation | AP / mIoU | Hurt |
| split features into 4 groups with separate offsets (ablation) | DySample · ADE20K (SegFormer-B1) | mIoU: 42.4 -> 43.2 | Improved |
| add dynamic scope factor (ablation) | DySample · ADE20K (SegFormer-B1) | mIoU: 43.2 -> 43.3 | Improved |
| use pixel shuffle + linear (PL) offset generation instead of linear + pixel shuffle (LP) | DySample · SegFormer / MaskFormer vs other models | mIoU | Mixed |
| replace bilinear upsampling with DySample-S+ in SegFormer-B1 | SegFormer-B1 · ADE20K | mIoU: 41.68 -> 43.58 | Improved |
| replace nearest with DySample-S+ in MaskFormer Swin-B | MaskFormer Swin-B · ADE20K | mIoU: 52.70 -> 53.91 | Improved |
| replace nearest with DySample+ in Faster R-CNN R50 | Faster R-CNN R50 · MS COCO | AP: 37.5 -> 38.7 | Improved |
| replace nearest with DySample+ in Panoptic FPN R50 | Panoptic FPN R50 · MS COCO | PQ: 40.2 -> 41.5 | Improved |
| use DySample (no guidance) instead of guided upsamplers FADE/SAPA for boundary quality | SegFormer-B1 · ADE20K | bIoU: 31.68 -> 29.93 | Hurt |
| replace bilinear with DySample+ in DepthFormer | DepthFormer Swin-T · NYU Depth V2 | delta < 1.25: 0.873 -> 0.878 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FADE — FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling (ECCV 2022) · better on 3 of 4 shared results
- CARAFE — CARAFE: Content-Aware ReAssembly of FEatures (ICCV 2019) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling (CVPR 2026 · similarity 0.88)
- FADE: A Task-Agnostic Upsampling Operator for Encoder–Decoder Architectures (IJCV 2024 · similarity 0.87)
- FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling (ECCV 2022 · similarity 0.86)
- CUF: Continuous Upsampling Filters (CVPR 2023 · similarity 0.85)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.84)
- RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation (CVPR 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.
