This repo is forked. The original readme contains many details about the package.

This readme contains instructions for retraining a model based on your data

Check out This Package for a wrapper to use this segmenter in ROS

I retrained on the YCB Video dataset.

1. download the dataset
2. Move (or create a symlink to) YCB_Video_Dataset in the data folder
3. prepare the dataset and training/test file lists by running prepare_ycb_video_dataset.py
4. Make your config file in the config directory. Choose your model, dataset, hyperparams, etc.
5. python train.py --cfg [your config file] --gpus [your gpu(s)]

To tailor the segmenter to your custom environment, I suggest, but have not tried:

1. Take pictures of your scene WITHOUT any ycb objects
2. Overlay the YCB objects from the data_syn portion of the YCB_Video_Dataset. These are YCB objects on a transparent background.
3. Include these at a 50/50 ratio with the original YCB Video dataset.

Alternatively, you could create your own dataset by segmenting the YCB objects from a bunch of pictures you take. That seems like a lot of work though....