Research map / Label-efficient and robust learning / Recognition and long-tail
GIFT: Learning Transformation-Invariant Dense Visual Descriptors via Group CNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 transformation-group features with group CNN embedding and bilinear pooling | GIFT · View-HP | PCK: 53.66 -> 67.15 | Helped |
| Use transformation-group features with group CNN embedding and bilinear pooling | GIFT · ES-HP | PCK: 12.16 -> 27.29 | Helped |
| Use transformation-group features with group CNN embedding and bilinear pooling | GIFT · ER-HP | PCK: 24.77 -> 48.93 | Helped |
| Fine-tune GIFT on GL3D | GIFT · MVS Average, DoG detector | rotation error (°): 2.189 -> 1.600 | Helped |
| Fine-tune GIFT on GL3D | GIFT · MVS Average, Superpoint detector | rotation error (°): 1.974 -> 1.490 | Helped |
| Add group convolution layers, increasing depth from 1 to 6 (ablation) | GIFT · HPSequence, ES-HP, and ER-HP | PCK: 59.61 -> 62.49; 63.71 -> 67.15; 21.74 -> 27.29; 39.68 -> 48.93 | Helped |
| Add group convolution layers, increasing depth from 3 to 6 (ablation) | GIFT · HPSequence, ES-HP, and ER-HP | PCK: 61.33 -> 62.49; 64.91 -> 67.15; 23.9 -> 27.29; 43.37 -> 48.93 | Helped |
| Replace bilinear pooling with average pooling (ablation) | GIFT-1 · Illum-HP, View-HP, ES-HP, and ER-HP | PCK: 59.61 vs 57.72; 63.71 vs 62.52; 21.74 vs 19.08; 39.68 vs 36.15 | Hurt |
| Replace bilinear pooling with max pooling (ablation) | GIFT-1 · Illum-HP, View-HP, ES-HP, and ER-HP | PCK: 54.31; 58.16; 19.37; 32.57 | Hurt |
| Replace bilinear pooling with subspace pooling (ablation) | GIFT-1 · Illum-HP, View-HP, ES-HP, and ER-HP | PCK: 47.21; 49.36; 14.85; 29.12 | Hurt |
| Use GIFT rather than vanilla CNN descriptors for dense matching | Dense correspondence estimation · HPSequence, ES-HP, and ER-HP | PCK: 26.96 -> 27.82; 32.92 -> 37.92; 4.64 -> 12.52; 14.02 -> 26.61 | Helped |
| Use direct group-feature comparison instead of group CNN embedding | GIFT-1 · Illum-HP | PCK: 59.61 -> 60.63 | Helped |
Closest papers by meaning
- Learning Rotation-Equivariant Features for Visual Correspondence (CVPR 2023 · similarity 0.88)
- Learning to Convolve: A Generalized Weight-Tying Approach (ICML 2019 · similarity 0.86)
- GPViT: A High Resolution Non-Hierarchical Vision Transformer with Group Propagation (ICLR 2023 · similarity 0.86)
- Auto-Encoding Transformations in Reparameterized Lie Groups for Unsupervised Learning (AAAI 2021 · similarity 0.85)
- Variational Partial Group Convolutions for Input-Aware Partial Equivariance of Rotations and Color-Shifts (ICML 2024 · similarity 0.84)
- Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis (NeurIPS 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.