Hang Zhao

This is a good question. Your points in 1. are exactly the trade-offs in this problem. A recent paper tries to solve this problem by sampling more patches around edges,...

The annotated images format are grayscale .png files

In our setup, label 0 is ignored during training. So if you have two classes, please set the labels as 1 and 2. @nanditam1

Hi, the cascade model was originally developed in Caffe, and now the performance is far from SOTA. So we did not try to reimplement it.

Check out the instance segmentation challenge results, our baseline results is 20.0 mAP. https://github.com/CSAILVision/placeschallenge/tree/master/instancesegmentation https://places-coco2017.github.io/

@loveis98 Our customized DataParallel only supports multi-gpu training. But in the upcoming version (```distributed``` branch), single-gpu training will be supported.

You might want to take a look at how data augmentation is done: https://github.com/CSAILVision/semantic-segmentation-pytorch/blob/master/mit_semseg/dataset.py#L110

Yes, it should be very straight forward if you write a script similar to ```eval_multipro.py```, which is for multi-gpu evaluation.

@scholltan Just added the script!

Augmentation only helps a little (