Research map / Detection and segmentation / CNN features
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, Hartwig Adam
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 the proposed decoder (ablation) | DeepLabv3 ResNet-101 OS16 · PASCAL VOC 2012 val | mIOU (%): 77.21% -> 78.85% | Improved |
| add the proposed decoder (ablation) | DeepLabv3 ResNet-101 eval OS8 · PASCAL VOC 2012 val | mIOU (%): 78.51% -> 79.35% | Improved |
| reduce low-level channels to 48 with 1x1 conv vs 8 (ablation) | decoder · PASCAL VOC 2012 val | mIOU (%): 77.61% -> 78.21% | Improved |
| use two 3x3 convs instead of one in decoder (ablation) | decoder · PASCAL VOC 2012 val | mIOU (%): 78.21% -> 78.85% | Improved |
| use three 3x3 convs instead of two (ablation) | decoder · PASCAL VOC 2012 val | mIOU (%): 78.85% -> 78.02% | Hurt |
| add Conv3 features to decoder (U-Net style) (ablation) | decoder · PASCAL VOC 2012 val | mIOU (%): 78.21% -> 78.61% | No clear effect |
| train with output stride 32 instead of 16 (ablation) | DeepLabv3+ ResNet-101 · PASCAL VOC 2012 val | mIOU (%): 78.85% -> 77.37% | Hurt |
| replace ResNet-101 with modified Xception (ablation) | DeepLabv3 OS16/16 · PASCAL VOC 2012 val | mIOU (%): 77.21% -> 79.17% | Improved |
| add decoder (ablation) | DeepLabv3 Xception OS16/16 · PASCAL VOC 2012 val | mIOU (%): 79.17% -> 79.93% | Improved |
| use depthwise separable conv in ASPP and decoder (ablation) | DeepLabv3+ Xception · PASCAL VOC 2012 val | Multiply-Adds: 89.76B -> 54.17B | No clear effect |
| pretrain on COCO (ablation) | DeepLabv3+ Xception SC · PASCAL VOC 2012 val | mIOU (%): 79.79% -> 82.20% | Improved |
| pretrain on JFT additionally (ablation) | DeepLabv3+ Xception SC COCO · PASCAL VOC 2012 val | mIOU (%): 82.20% -> 83.03% | Improved |
Closest papers by meaning
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · similarity 0.91)
- Dense Decoder Shortcut Connections for Single-Pass Semantic Segmentation (CVPR 2018 · similarity 0.88)
- A Holistically-Guided Decoder for Deep Representation Learning With Applications to Semantic Segmentation and Object Detection (TPAMI 2021 · similarity 0.88)
- EfficientFCN: Holistically-guided Decoding for Semantic Segmentation (ECCV 2020 · similarity 0.88)
- Decoders Matter for Semantic Segmentation: Data-Dependent Decoding Enables Flexible Feature Aggregation (CVPR 2019 · similarity 0.87)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.85)
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.
