Research map / Detection and segmentation / CNN features
Scattering Vision Transformer: Spectral Mixing Matters
Badri Narayana Patro, Vijay Srinivas Agneeswaran
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 DTCWT scattering with low-frequency tensor mixing and high-frequency Einstein mixing (ablation) | SVT · ImageNet-1K | Top-1 accuracy: 83.70 -> 84.20 | Improved |
| Place scatter layers first, followed by attention layers (ablation) | SVT-H-S · ImageNet-1K | Top-1 accuracy: 83.1 -> 84.2 | Improved |
| Use initial scatter layers rather than initial convolutional layers (ablation) | SVT-H-S · ImageNet-1K | Top-1 accuracy: 84.0 -> 84.2 | Improved |
| Replace DTCWT scattering with Fourier or wavelet spectral alternatives (ablation) | SVT · ImageNet-1K | Top-1 accuracy: 83.1; 84.02; 84.09; 83.70 -> 84.20 | Hurt |
| Use tensor mixing in both frequency components instead of tensor-low/Einstein-high mixing (ablation) | SVT · ImageNet-1K | Top-1 accuracy: 83.97 -> 84.20 | Hurt |
| Use Einstein mixing in low frequency and tensor mixing in high frequency (ablation) | SVT · ImageNet-1K | Top-1 accuracy: 83.87 -> 84.20 | Hurt |
| Remove high-frequency channel mixing (ablation) | SVT_TTEE · ImageNet-1K | Top-1 accuracy: 84.06 -> 84.20 | Hurt |
| Remove high-frequency token mixing (ablation) | SVT_TTEE · ImageNet-1K | Top-1 accuracy: 84.12 -> 84.20 | Hurt |
| Use all six directional orientations rather than fewer orientations (ablation) | SVT-H-S · ImageNet-1K | Top-1 accuracy: 83.2; 83.4; 83.7 -> 84.2 | Improved |
| Use SVT-H-B instead of DeiT-B/16 | ImageNet-pretrained transfer learning · CIFAR-100 | Top-1 accuracy: 90.8 -> 91.2 | Improved |
| Use SVT-H-S instead of PVTv2-B2 as the Mask R-CNN backbone | Mask R-CNN · COCO val2017 | APᵇ: 45.3 -> 46.0 | Improved |
| Use SVT-H-S instead of PVTv2-B2 as the Mask R-CNN backbone | Mask R-CNN · COCO val2017 | APᵐ: 41.2 -> 41.9 | Improved |
Closest papers by meaning
- Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning (ECCV 2022 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.86)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.85)
- Visual Fourier Prompt Tuning (NeurIPS 2024 · similarity 0.85)
- Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · similarity 0.85)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.85)
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.
