Research map / Detection and segmentation / CNN features
TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba
Xiaowen Ma, Zhenliang Ni, Xinghao Chen
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 Mamba global modeling to a conv-only hybrid (ablation) | TinyViM-S layout · ImageNet-1K val (Mamba input ablation, TinyViM-S layout) | top-1 accuracy (%): 77.5 -> 79.1 | Improved |
| feed only low-frequency components to Mamba instead of full features (ablation) | TinyViM-S layout · ImageNet-1K val (Mamba input ablation, TinyViM-S layout) | top-1 accuracy (%): 79.1 -> 79.0 | No clear effect |
| feed only low-frequency components to Mamba instead of full features (ablation) | TinyViM-S layout · ImageNet-1K val (Mamba input ablation, TinyViM-S layout) | throughput (im/s): 1673 -> 2574 | Improved |
| feed only high-frequency input to Mamba instead of low only (ablation) | TinyViM-S layout · ImageNet-1K val (Mamba input ablation, TinyViM-S layout) | top-1 accuracy (%): 79.0 -> 78.6 | Hurt |
| add high-frequency input in parallel to Mamba on top of low (ablation) | TinyViM-S layout · ImageNet-1K val (Mamba input ablation, TinyViM-S layout) | throughput (im/s): 2574 -> 1509 | Hurt |
| add frequency ramp (alpha 0.25/0.5/0.5/0.75) to inception (ablation) | TinyViM-S · ImageNet-1K val | top-1 accuracy (%): 79.0 -> 79.2 | Improved |
| add frequency ramp inception versus decoupling all channels (ablation) | TinyViM-S · ImageNet-1K val | throughput (im/s): 2478 -> 2563 | Improved |
| use ramp alpha increasing with depth instead of inverse ramp (ablation) | TinyViM-S · ImageNet-1K val | top-1 accuracy (%): 78.8 -> 79.2 | Improved |
| use axial depthwise kernel 7 instead of 3x3 depthwise (ablation) | TinyViM-S · ImageNet-1K val | top-1 accuracy (%): 79.0 -> 79.2 | Improved |
| increase axial kernel from 7 to 9 (ablation) | TinyViM-S · ImageNet-1K val | top-1 accuracy (%): 79.2 -> 79.2 | No clear effect |
| add re-parameterization (ablation) | TinyViM-L · ImageNet-1K val | top-1 accuracy (%): 83.1 -> 83.3 | Improved |
| add re-parameterization (ablation) | TinyViM-S · ImageNet-1K val | top-1 accuracy (%): 79.2 -> 79.2 | No clear effect |
Closest papers by meaning
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.88)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.87)
- MambaVision: A Hybrid Mamba-Transformer Vision Backbone (CVPR 2025 · similarity 0.87)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.87)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.86)
- DiMSUM: Diffusion Mamba - A Scalable and Unified Spatial-Frequency Method for Image Generation (NeurIPS 2024 · similarity 0.86)
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.
