Zilong Huang
Zilong Huang
@Raven1327 Do you run this repo on PASCAL VOC dataset or another dataset? You can simply modify the NUM_CLASSES in [this line](https://github.com/speedinghzl/Pytorch-Deeplab/blob/master/train.py#L33) for your dataset.
@mcever Have you solved the problem? You throw much information which may be irrelevant to your question. Maybe you can check your caffe by running `make runtest` before running DSRG.
Yes，I keep changing the statistics of BN layers during training. This is very old code for semantic segmentation. Some details may be not maintained correctly. In fact, it's helpful to...
谢谢你的关注。可以通过CAM得到person的seed，可能loss需要换成sigmoid+cross entropy loss。
1000+图训练分类器是可以的（只有最后一层参数是随机初始化的）。 另外1000+图片都是有人的吗？如果是可能数据集并不适合训练CAM来定位人的位置。
如果可以把人去掉，意味着可以生成pixel-level label，就没必要使用弱监督分割了。直接使用pixel-level label训练一个人的segmentation network会比弱监督分割的结果要好。
Maybe you should set align_corners=True for Upsample for Pytroch > 0.4.
Do you evaluate the VOC_scenes_20000.pth I provided?
@WilliamLwj Thanks for your feedback and your produced results. I'm not sure whether the performance gap is caused by the different Pytorch versions (0.2 vs 1.0).
The baseline does not include expand loss. The training hyperparameters are the same as DSRG (in this repo). The [cues](https://drive.google.com/open?id=1cHyhjul9srPlwcl4xqrYR9MwzhFGwKXU) have been provided in this repo.