domain-adaptation-for-segmentation-toolbox!

domain adaptation for segmentation toolbox(Pytorch) - deeplabev2, adaptsegnet...etc

This code is heavily bor

rowed from Pytorch-Deeplab.

This code is heavily borrowed from Adaptsegnet.

Dataset

• Download the GTA5 Dataset as the source domain.
• Download the Cityscapes Dataset as the target domain.

Data\
    Cityscapes\
        gtFine\
        gtFine_trainvaltest\
        leftImg8bit\
        meta\
    GTA5\
        images\
        labels\

Training and Testing script

GTA Only

python train_GTA_only.py --snapshot-dir ./snapshots/GTA2Cityscapes_single \
                                     --lambda-seg 0.0 \
                                     --lambda-adv-target1 0.0 --lambda-adv-target2 0.001


python evaluate_cityscapes.py --restore-from ./snapshots/GTA2Cityscapes_single/GTA5_95000.pth


python compute_iou.py ./data/Cityscapes/gtFine/val result/cityscapes

AdaptSegNet

python train_gta2cityscapes_multi.py --snapshot-dir ./snapshots/GTA2Cityscapes_single \
                                     --lambda-seg 0.0 \
                                     --lambda-adv-target1 0.0 --lambda-adv-target2 0.001


python evaluate_cityscapes.py --restore-from ./snapshots/GTA2Cityscapes_single/GTA5_115000.pth


python compute_iou.py ./data/Cityscapes/gtFine/val result/cityscapes

Cityscapes Only

python train_cityscapes_only.py

python evaluate_cityscapes.py --restore-from ./snapshots/GTA2Cityscapes_single/Cityscapes_95000.pth

python compute_iou.py ./data/Cityscapes/gtFine/val result/cityscapes

Reslut

Method	road	sidewalk	building	wall	fence	pole	light	sign	vegetation	terrain	sky	person	rider	car	truck	bus	train	motocycle	bicycle	mIoU
DeepLabV2-GTA5	85.28	20.23	69.0	21.0	14.13	22.36	31.83	15.74	65.21	19.79	68.5	55.28	26.24	72.13	25.74	32.48	1.34	29.35	38.11	37.57
AdaptSegNet(patch)	86.49	24.52	81.14	24.59	23.74	29.02	35.7	25.43	83.46	33.03	76.06	57.88	29.41	78.87	30.45	26.57	2.74	28.89	19.08	41.95
AdaptSegNet(patch, image)	88.23	26.70	84.62	28.09	25.22	31.29	38.9	27.12	89.49	35.23	79.99	59.08	31.20	80.12	32.60	29.12	5.92	30.11	22.91	43.21
AdaptSegNet(patch)+image sampling	88.00	25.90	86.23	29.19	27.42	32.21	39.92	28.35	88.99	36.42	80.47	60.24	32.10	80.09	34.94	30.54	4.99	34.52	24.09	44.89
CCM	87.20	28.44	86.59	24.12	24.59	32.48	39.32	27.34	90.91	32.41	82.99	60.32	32.10	85.23	30.44	29.51	3.99	30.21	22.49	42.80
DeepLabV2-cityscapes	96.29	75.58	87.79	38.07	39.63	43.46	46.63	62.81	88.24	52.41	89.53	69.73	49.5	91.49	66.23	69.76	45.01	49.08	65.16	64.55