Research map / Detection and segmentation / CNN features
Beyond MACs: Hardware Efficient Architecture Design for Vision Backbones
Moritz Nottebaum, Matteo Dunnhofer, Christian Micheloni
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 fused MBConv in place of unfused MBConv (ablation) | LowFormer-B1 · ImageNet-1K | top-1 accuracy: 79.1 -> 79.9 | Improved |
| add fused MBConv in place of unfused MBConv (ablation) | LowFormer-B1 · ImageNet-1K | GPU throughput: 3558 -> 4237 | Improved |
| add Lowtention in place of ReLU linear attention (ablation) | LowFormer-B1 · ImageNet-1K | top-1 accuracy: 79.6 -> 79.9 | Improved |
| add downscaling of feature maps in Lowtention (ablation) | LowFormer-B1 · ImageNet-1K | top-1 accuracy: 79.9 -> 79.9 | No clear effect |
| add downscaling of feature maps in Lowtention (ablation) | LowFormer-B1 · ImageNet-1K | TX2 latency: 13.4 -> 11.7 | Improved |
| add channel compression in Lowtention (ablation) | LowFormer-B1 · ImageNet-1K | GPU throughput: 3921 -> 4237 | Improved |
| add Lowtention in place of original MHSA (ablation) | LowFormer-B1 · ImageNet-1K | top-1 accuracy: 79.8 -> 79.9 | Improved |
| add Lowtention in place of original MHSA at 1024x1024 input (ablation) | LowFormer-B1 · ImageNet-1K at 1024x1024 | TX2 latency: 220% higher latency without downsampling | Improved |
| remove MLP from LowFormer-B1 (ablation) | LowFormer-B1 (B1_mlpless) · ImageNet-1K | top-1 accuracy: 79.9 -> 78.8 | Hurt |
| remove MLP and reduce depth (shallow) (ablation) | LowFormer-B1 · ImageNet-1K | top-1 accuracy: 78.8 -> 77.2 | Hurt |
| remove Lowtention (conv only) from B1.5 shallow (ablation) | LowFormer-B1.5 · ImageNet-1K | top-1 accuracy: 79.7 -> 78.8 | Hurt |
| remove MLP from LowFormer-B3 (edge variant E3) | LowFormer-B3 · ImageNet-1K | TX2 latency: 32.5 -> 25.0 | Improved |
Closest papers by meaning
- Battle of the Backbones: A Large-Scale Comparison of Pretrained Models across Computer Vision Tasks (NeurIPS 2023 · similarity 0.88)
- RIFormer: Keep Your Vision Backbone Effective but Removing Token Mixer (CVPR 2023 · similarity 0.86)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.85)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.84)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.84)
- FreeNet: Liberating Depth-Wise Separable Operations for Building Faster Mobile Vision Architectures (AAAI 2025 · 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.
