DeepLab-V1-PyTorch

Code for ICLR 2015 deeplab-v1 paper "Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs", backbone is deeplab-largeFOV.

Config

python 3.7 / pytorch 1.2.0
pydensecrf
opencv

Datasets

Pascal VOC 2012 Dataset
- extract 'VOCtrainval_11-May-2012.tar' to 'VOCdevkit/'
Pascal VOC 2012 Augment Dataset
- I have download it, 'VOCdevkit/SegmentationClassAug.zip', you show extract it to 'VOCdevkit/VOC2012/ Finally, it should like this

/VOCdevkit
└── VOC2012
    ├── Annotations
    ├── ImageSets
    ├── JPEGImages
    ├── SegmentationObject
    └── SegmentationClass
    └── SegmentationClassAug

Performance

Compore with paper

Train set	Eval set	CRF?	Code	Pixel Accuracy	Mean Accuracy	Mean IoU	FreqW IoU
trainaug	val		DeepLab-LargeFOV	-	-	62.25	-
			Ours-step	90.20	80.92	62.34	83.58
		✓	DeepLab-CRF-LargeFOV	-	-	67.64	-
		✓	Ours-step	92.23	79.99	67.58	86.32

e.g. We use ImageNet-1000 to train a classification task for backbone DeepLab-LargeFOV（VGG16）。

Learning rate policy

Train set	Eval set	CRF?	Code	Pixel Accuracy	Mean Accuracy	Mean IoU	FreqW IoU
trainaug	val		Ours-step	90.20	80.92	62.34	83.58
			Ours-poly	91.24	80.29	64.838	85.00
			Ours-poly2	91.35	81.61	65.44	85.26
		✓	Ours-step	92.23	79.99	67.58	86.32
			Ours-poly	92.65	79.92	68.72	86.98
			Ours-poly2	92.89	80.39	69.60	87.35

e.g. The learning rate policy of "step" and "poly", mentioned by [deeplab v2](http://arxiv.org/pdf/1606.00915.pdf). Ours-poly2 means to add scale method to dataset augment.

Usage

Train

python main.py --type=train

Test

without CRF

python main.py --type=test

with CRF

python main.py --type=test --use_crf

Evaluate

python evalate.py

References

Liang-Chieh Chen*, George Papandreou*, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. (*equal contribution). Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs. ICLR, 2015.
Project / Code / arXiv paper
deeplab-v2-pytorch

FMsunyh/DeepLab-V1-PyTorch