Research map / Detection and segmentation / CNN features
LiFT: A Surprisingly Simple Lightweight Feature Transform for Dense ViT Descriptors
Saksham Suri, Matthew Walmer, Kamal Gupta, Abhinav Shrivastava
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 |
|---|---|---|---|
| add LiFT feature transform to DINO features | DINO ViT-S/16 · SPair-71k (224 res) | PCK@0.1: 24.76 -> 28.68 | Improved |
| add LiFT to DINO instead of using a larger ViT-B backbone | DINO ViT-S/16 · SPair-71k (224 res) | PCK@0.1: 28.68 vs 24.90 (ViT-B/16) | Improved |
| add LiFT to DINO | DINO ViT-S/16 · DAVIS (224 res) | J&F mean: 33.0 -> 44.3 | Improved |
| add LiFT to DINO | DINO ViT-S/16 · COCO20K object discovery (56 res) | CorLoc: 16.28 -> 27.72 | Improved |
| add LiFT to ViTDet Mask R-CNN | ViTDet(MR) · COCO detection | AP: 39.50 -> 45.98 | Improved |
| add LiFT to ViTDet Cascade R-CNN | ViTDet(CR) · COCO segmentation | AP: 40.93 -> 41.42 | Improved |
| train LiFT with self-supervision versus random initialization (ablation) | DINO+LiFT · SPair-71k (224 res) | PCK@0.1: 4.21 -> 28.68 | Improved |
| add image input to LiFT (ablation) | DINO+LiFT · SPair-71k (224 res) | PCK@0.05: 13.28 -> 14.72 | Improved |
| use cosine distance rather than L1 or L2 reconstruction loss (ablation) | DINO+LiFT · SPair-71k (224 res) | PCK@0.1: 27.77 (L1), 28.17 (L2) -> 28.68 | Improved |
| train LiFT longer (5 to 100 epochs) (ablation) | DINO+LiFT · SPair-71k (224 res) | PCK@0.1: 28.45 -> 29.11 | No clear effect |
| apply LiFT recursively (2xLiFT) (ablation) | DINO+LiFT · SPair-71k (448 res) | PCK@0.1: 28.68 -> 29.45 at 224; 31.38 -> 31.35 at 448 | Mixed |
| apply LiFT trained on a different backbone (ablation) | DINO+LiFT · SPair-71k (448 res) | PCK@0.1: 28.6 -> 9.86 (MoCo-trained LiFT on DINO would be 16.02) | Hurt |
Closest papers by meaning
- AiluRus: A Scalable ViT Framework for Dense Prediction (NeurIPS 2023 · similarity 0.87)
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.86)
- LIFT: Learned Invariant Feature Transform (ECCV 2016 · similarity 0.86)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.85)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.85)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 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.
