Research map / Architectures and efficiency / Attention and transformers
Axial-DeepLab: Stand-Alone Axial-Attention for Panoptic Segmentation
Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, Liang-Chieh 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 batch normalization to attention layers (ablation) | Conv-Stem + Attention · ImageNet val | top-1 (%): 77.4 -> 77.7 | Improved |
| add position-sensitive self-attention (ablation) | Conv-Stem + Attention · ImageNet val | top-1 (%): 77.7 -> 78.1 | Improved |
| factorize into axial-attention instead of PS-Attention (conv stem) (ablation) | Conv-Stem PS-Attention · ImageNet val | top-1 (%): 78.1 -> 77.5 | No clear effect |
| use Full Axial-Attention instead of Full Attention (ablation) | Full Attention · ImageNet val | top-1 (%): 77.6 -> 78.1 | Improved |
| add ASPP (ablation) | ResNet-50 backbone · Cityscapes val | PQ (%): 58.1 -> 59.8 | Improved |
| add ASPP (ablation) | PS-Attention backbone · Cityscapes val | PQ (%): 59.9 -> 61.5 | Improved |
| add position-sensitivity to attention (ablation) | Attention backbone · Cityscapes val | PQ (%): 58.7 -> 59.9 | Improved |
| add axial-attention instead of PS-Attention, no ASPP (ablation) | PS-Attention · Cityscapes val | PQ (%): 59.9 -> 62.6 | Improved |
| increase axial span from 5x5 to 65x65 (ablation) | Axial-ResNet-L · Cityscapes val | PQ (%): 59.1 -> 64.2 | Improved |
| increase output stride from 16 to 32 (ablation) | Axial-ResNet-L · Cityscapes val | PQ (%): 63.9 -> 63.9 | No clear effect |
| replace dual convolutional decoders with axial-decoder (ablation) | Axial-ResNet-L OS16 · Cityscapes val | PQ (%): 63.9 -> 63.7 | No clear effect |
| use multi-scale inference | Axial-DeepLab-L · COCO val | PQ (%): 43.4 -> 43.9 | Improved |
Closest papers by meaning
- Channelized Axial Attention – considering Channel Relation within Spatial Attention for Semantic Segmentation (AAAI 2022 · similarity 0.87)
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · similarity 0.86)
- Explicitly Modeled Attention Maps for Image Classification (AAAI 2021 · similarity 0.86)
- Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation (CVPR 2020 · similarity 0.86)
- MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers (CVPR 2021 · similarity 0.86)
- Stand-Alone Self-Attention in Vision Models (NeurIPS 2019 · 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.
