The project is an reimplementation of DeepLabV2-ResNet in Pytorch for semantic image segmentation on the PASCAL VOC dataset.

Attention:

• This proj is based on pytorch-deeplab-resnet.
• In this proj, we change the loss-calculate method, which ignores the background labels.
• Evaluation of a single-scale model on the PASCAL VOC validation dataset leads to 74.95% mIoU VOC12_50000.pth which is almost equal to 75.1 reimplemented by DrSleep.

• The running means and variances of batch normalization layer of ResNet will be updated. I will try to use for i in self.bn.parameters(): i.requires_grad = Falsefor ResNet layer to verify the performance.
• Pytorch is more flexible to use multi-gpu than TensorFlow, just usetorch.nn.DataParallel(model).cuda(). But for BatchNorm synchronization across multipe GPUs I will try it later.

1. python 3
2. pytorch 0.3.1
3. numpy
4. opencv
5. pillow

1. Download this proj git clone https://github.com/CarryJzzZ/pithy-conky-colors.gitand enter it.
2. Download the init.pth which contains MS COCO trained weights. init.pth and put it into dataset folder.
3. Change DATA_DIRECTORYline 24 of train.py to VOC2012 where you store the pascal voc12 dataset. (trainning dataset is based on SBD)
4. run python train.py --random-mirror --random-scale --gpu 0

1. change RESTORE_FROM of evaluate.py to your trained .pth file or you can download demo weights
2. run python evaluate.py
3. predictions are stored in outputs