Research map / Architectures and efficiency / CNN design
TVConv: Efficient Translation Variant Convolution for Layout-Aware Visual Processing
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 |
|---|---|---|---|
| Replace depthwise convolution with TVConv | MobileNetV2 on LFW · LFW | computational cost: 3.05x | Helped |
| Use TVConv in MBx0.5 rather than the original MBx1.0 | MobileNetV2 on LFW · LFW | throughput: 2.3x | Helped |
| Replace depthwise convolution with TVConv | MobileNetV2 MBx0.1 · Face verification datasets | mean accuracy: 4.21% | Helped |
| Use TVConv instead of the depthwise-convolution variant | DeepLabV3+ optic disc/cup segmentation · Unseen domains, mean | DSC: 86.10 -> 87.31 | Helped |
| Use TVConv instead of the depthwise-convolution variant | DeepLabV3+ optic disc/cup segmentation · Unseen domains, mean | HD: 26.26 -> 24.32 | Helped |
| Initialize affinity maps from data statistics rather than constant 1 (ablation) | TVConv · LFW, MBx0.1/0.5/1.0 | error (%): 2.84 ± 0.11 -> 2.92 ± 0.22; 1.73 ± 0.07 -> 1.75 ± 0.18; 1.43 ± 0.12 -> 1.62 ± 0.20 | Hurt |
| Increase weight-generating block width to 128 channels from the default 64 (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.60 ± 0.10 | Hurt |
| Reduce weight-generating block width from 64 to 32 or 8 channels (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.65 ± 0.21; 2.31 ± 0.12 -> 2.65 ± 0.12 | Hurt |
| Increase weight-generating block depth from 3 to 4 layers (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.49 ± 0.16 | Hurt |
| Reduce weight-generating block depth from 3 to 2 or 1 layer (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.44 ± 0.15; 2.31 ± 0.12 -> 2.48 ± 0.27 | Mixed |
| Increase affinity-map channels from 4 to 8 (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.49 ± 0.16 | Hurt |
| Reduce affinity-map channels from 4 to 2 or 1 (ablation) | TVConv MBx0.2 · LFW | error (%): 2.31 ± 0.12 -> 2.44 ± 0.19; 2.31 ± 0.12 -> 2.52 ± 0.15 | Hurt |
Closest papers by meaning
- BA-SAM: Scalable Bias-Mode Attention Mask for Segment Anything Model (CVPR 2024 · similarity 0.85)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.85)
- Vitron: A Unified Pixel-level Vision LLM for Understanding, Generating, Segmenting, Editing (NeurIPS 2024 · similarity 0.84)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.84)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 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.