Research map / Detection and segmentation / CNN features
LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models
Haiwen Huang, Anpei Chen, Volodymyr Havrylov, Andreas Geiger, Dan Zhang
This paper tried a common change: Add positional encoding
add sinusoidal positional encoding and 3x3 image conv · Cityscapes segmentation (Stage 1, 50k) · mIoU: 43.13 -> 48.63 Improved
What 646 accepted papers report for this change
- 525 improved
- 58 no clear effect
- 7 mixed
- 56 hurt
- Learning-Augmented Scalable Linear Assignment Problem Optimization via Neural Dual Warm-Starts ICML 2026 · add positional encodings to features (d=13 -> d=21) · Block-structured, N=4096 · speedupImproved
- Mesh Based Simulations with Spatial and Temporal awareness ICML 2026 · use learned positional embedding, learnable relative bias or weighted adjacency instead of RoPE · Aneurysm · RMSEHurt
- Sparse ActionGen: Accelerating Diffusion Policy with Real-time Pruning ICML 2026 · add sinusoidal positional encoding to the pruner · Robomimic Tool Hang · success rate (%): 28 -> 54Improved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 |
|---|---|---|---|
| add Stage 2 self-distillation pseudo-GT on top of Stage 1 mask-bicubic (ablation) | LoftUp · Cityscapes segmentation | IoU: 50.43 -> 53.10 | Improved |
| use SAM mask-based pseudo-GT instead of LiFT 2x features (ablation) | LoftUp · Cityscapes segmentation | IoU: 35.85 -> 50.43 | Improved |
| use full-resolution pseudo-GT instead of per-image optimized features (ablation) | LoftUp · COCO-Stuff segmentation | IoU: 59.08 -> 61.11 | Improved |
| use coordinate-based cross-attention architecture instead of LIIF (ablation) | LoftUp · COCO-Stuff segmentation | mIoU: 52.24 -> 61.11 | Improved |
| apply LoftUp pseudo-GT objective to other upsamplers (ablation) | resize-conv / FeatUp-JBU · COCO-Stuff segmentation | mIoU: 56.05 -> 56.05 (resize-conv with LoftUp objective); FeatUp 56.30 -> 57.90 | Improved |
| add sinusoidal positional encoding and 3x3 image conv (ablation) | LoftUp · Cityscapes segmentation (Stage 1, 50k) | mIoU: 43.13 -> 48.63 | Improved |
| increase cross-attention blocks from 2 to 3 (ablation) | LoftUp · COCO-Stuff segmentation (Stage 1) | mIoU: 58.65 -> 58.35 | No clear effect |
| increase training set from 50k to 1M images (ablation) | LoftUp · COCO-Stuff segmentation (Stage 1) | mIoU: 58.65 -> 59.87 | Improved |
| upsample 224-input features to 448 resolution instead of 224 (ablation) | LoftUp · COCO-Stuff segmentation | mIoU: 61.11 -> 60.61 | Hurt |
| replace 4x larger VFM input with LoftUp upsampling at 224 input | DINOv2-S/14 · COCO-Stuff segmentation | mIoU: 56.79 -> 61.11 | Improved |
| apply LiFT upsampler to VFM for interactive segmentation | DINOv2-S/14 · GrabCut interactive segmentation | IoU@1 click: 64.90 -> 29.66 | Hurt |
Later papers that compared with this method
2 later accepted papers report LoftUp as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- UPLiFT: Efficient Pixel-Dense Feature Upsampling with Local Attenders (CVPR 2026 · better on 4 of 4 shared results)
- Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling (CVPR 2026 · better on 4 of 4 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LiFT — LIFT: Learning 4D LiDAR Image Fusion Transformer for 3D Object Detection (CVPR 2022) · better on 5 of 5 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
- JAFAR: Jack up Any Feature at Any Resolution (NeurIPS 2025 · similarity 0.87)
- FeatUp: A Model-Agnostic Framework for Features at Any Resolution (ICLR 2024 · similarity 0.85)
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.84)
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · similarity 0.83)
- Coarse-To-Fine Tensor Trains for Compact Visual Representations (ICML 2024 · similarity 0.83)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 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.
