Research map / Architectures and efficiency / Attention and transformers
Inverse Compositional Spatial Transformer Networks
Check 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 |
|---|---|---|---|
| Reuse the geometric predictor across recurrent warp operations rather than use one warp predictor (ablation) | Planar image alignment · Planar image alignment, σ=2.5–10 | test error: 2.699 -> 0.027; 5.576 -> 0.125; 9.491 -> 0.245; 9.218 -> 1.305 | Helped |
| Add recurrent warp operations with a shared predictor (ablation) | MNIST architecture (a) · MNIST perturbed test set | test error: 3.687 % -> 1.230 % | Helped |
| Add recurrent warp operations with a shared predictor (ablation) | MNIST architecture (b) · MNIST perturbed test set | test error: 8.545 % -> 1.703 % | Helped |
| Add recurrent warp operations with a shared predictor (ablation) | GTSRB · GTSRB perturbed test set | test error: 5.011 % -> 3.184 % | Helped |
| Replace the original single STN with recurrent IC-STN alignment | MNIST architecture (a) · MNIST perturbed test set | test error: 4.944 % -> 1.230 % | Helped |
| Replace the original single STN with recurrent IC-STN alignment | MNIST architecture (b) · MNIST perturbed test set | test error: 9.325 % -> 1.703 % | Helped |
| Replace the original single STN with recurrent IC-STN alignment | GTSRB · GTSRB perturbed test set | test error: 6.495 % -> 3.184 % | Helped |
| Increase the number of shared recurrent warp operations (ablation) | IC-STN · MNIST perturbed test set, architecture (a) | test error: 1.905 % -> 1.230 % | Helped |
| Increase the number of shared recurrent warp operations (ablation) | IC-STN · MNIST perturbed test set, architecture (b) | test error: 3.717 % -> 1.703 % | Helped |
| Increase the number of shared recurrent warp operations (ablation) | IC-STN · GTSRB perturbed test set | test error: 4.122 % -> 3.184 % | Helped |
| Use four separate compositional transformers instead of a shared recurrent predictor (ablation) | MNIST architecture (a) · MNIST perturbed test set | test error: 1.476 % -> 1.230 % | Helped |
| Use recurrent alignment instead of a CNN without spatial transformations | CNN · MNIST perturbed test set, architecture (a) | test error: 6.597 % -> 1.230 % | Helped |
Closest papers by meaning
- Hide & Seek: Transformer Symmetries Obscure Sharpness & Riemannian Geometry Finds It (ICML 2025 · similarity 0.83)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.83)
- SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers (CVPR 2025 · similarity 0.83)
- Compositional Inversion for Stable Diffusion Models (AAAI 2024 · similarity 0.82)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.82)
- Transformer Fusion with Optimal Transport (ICLR 2024 · similarity 0.82)
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.