Research map / Detection and segmentation / CNN features
Dual-stream Network for Visual Recognition
Read the paperCheck 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 |
|---|---|---|---|
| Use dual local/global streams with DS-Net-S* rather than DeiT-S | Image classification · ImageNet-1k validation | Top-1 (%): +2.4% | Helped |
| Add Inter-scale Alignment to DS-Net-S (ablation) | DS-Net-S · ImageNet-1k validation | Top-1 (%): 81.9 -> 82.3 | Helped |
| Use a 50/50 global-local channel split (ablation) | DS-Net-T · ImageNet-1k validation | Top-1 (%): 77.1 -> 78.1 | Helped |
| Remove local feature propagation fL (ablation) | DS-Net-T* · ImageNet-1k validation | Top-1 (%): 79.0 -> 76.7 | Hurt |
| Remove global feature propagation fG (ablation) | DS-Net-T* · ImageNet-1k validation | Top-1 (%): 79.0 -> 76.6 | Hurt |
| Remove global-to-local co-attention (ablation) | DS-Net-T* · ImageNet-1k validation | Top-1 (%): 79.0 -> 76.5 | Hurt |
| Remove local-to-global co-attention (ablation) | DS-Net-T* · ImageNet-1k validation | Top-1 (%): 79.0 -> 76.4 | Hurt |
| Remove both directions of Inter-scale Alignment (ablation) | DS-Net-T* · ImageNet-1k validation | Top-1 (%): 79.0 -> 78.1 | Hurt |
| Replace FPN with DS-FPN | DS-Net-T with RetinaNet · MSCOCO 2017 | AP bbox: 39.0 -> 40.4 | Helped |
| Insert DS-Blocks in lateral connections and extra feature outputs (ablation) | DS-Net-T with RetinaNet FPN · MSCOCO RetinaNet | AP bbox: 39.0 -> 40.4 | Helped |
| Add DS-Net-S* as backbone instead of ResNet-50 | RetinaNet · MSCOCO minival | AP bbox: 36.3 -> 42.7 | Helped |
| Add DS-Net-S* as backbone instead of ResNet-50 | Mask R-CNN · MSCOCO minival | AP segm: 34.7 -> 40.2 | Helped |
Closest papers by meaning
- Single Image Reflection Separation via Dual-Stream Interactive Transformers (NeurIPS 2024 · similarity 0.86)
- Learning Visual Abstract Reasoning through Dual-Stream Networks (AAAI 2024 · similarity 0.85)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.84)
- Scaling White-Box Transformers for Vision (NeurIPS 2024 · similarity 0.84)
- Transformer Fusion with Optimal Transport (ICLR 2024 · similarity 0.84)
- Harnessing Edge Information for Improved Robustness in Vision Transformers (AAAI 2024 · 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.