Research map / Architectures and efficiency / Neural architecture search
SparseMask: Differentiable Connectivity Learning for Dense Image Prediction
Huikai Wu, Junge Zhang, Kaiqi Huang
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 batch normalization before the connection weights (ablation) | SparseMask · PASCAL VOC 2012 val | mIoU: 72.21% -> 73.18% | Improved |
| add the binary part of the sparse loss to the task loss (ablation) | FDN search · PASCAL VOC 2012 val | mIoU: 72.72% -> 73.23% | Improved |
| add the sparse part (L_sp) on top of L_task + L_bp (ablation) | SparseMask · PASCAL VOC 2012 val | mIoU / parameters: 73.23% -> 73.18% | No clear effect |
| add the sparse part alone to the task loss (ablation) | FDN search · PASCAL VOC 2012 val | mIoU: 72.72% -> 72.04% | Hurt |
| replace the dense FDN with the searched sparse connectivity | SparseMask · PASCAL VOC 2012 val | FPS: 37 -> 98 | Improved |
| replace U-Net with the searched connectivity | MobileNet-V2 decoder · PASCAL VOC 2012 val | mIoU: 64.72% -> 73.18% | Improved |
| replace SparseACN with the searched connectivity | MobileNet-V2 decoder · PASCAL VOC 2012 val | mIoU: 72.23% -> 73.18% | Improved |
| replace FCN with the searched connectivity | MobileNet-V2 decoder · PASCAL VOC 2012 val | mIoU: 63.80% -> 73.18% | Improved |
| replace Deeplab-V3 decoder with the searched connectivity | MobileNet-V2 · PASCAL VOC 2012 val | mIoU / FPS: 72.51% -> 73.18%; 50 -> 98 FPS | Improved |
| replace the proposed sparse loss with L1 loss (ablation) | SparseMask search · PASCAL VOC 2012 val | FPS / mIoU: nearly 3 times slower at 1e-3 | Hurt |
| increase L1 loss strength from 1e-3 to 1e-2 (ablation) | L1-based search · PASCAL VOC 2012 val | mIoU / FPS | Mixed |
| increase MaskConnect k from 2 to 5 (ablation) | MaskConnect search · PASCAL VOC 2012 val | mIoU | Improved |
Closest papers by meaning
- DCNAS: Densely Connected Neural Architecture Search for Semantic Image Segmentation (CVPR 2021 · similarity 0.88)
- MaskConnect: Connectivity Learning by Gradient Descent (ECCV 2018 · similarity 0.88)
- DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTs (ECCV 2024 · similarity 0.87)
- Convolutional Networks with Dense Connectivity (TPAMI 2019 · similarity 0.87)
- Densely Connected Convolutional Networks (CVPR 2017 · similarity 0.86)
- Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness (ICLR 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.
