search

facebookresearch / maskrcnn-benchmark

Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.
MIT License
9.27k stars 2.5k forks source link

Add data agumentation in config #143

Open guanbin1994 opened 5 years ago

guanbin1994 commented 5 years ago

❓ Questions and Help

My own dataset is small, so i need data agumentation like flip etc. What should i do? Can i only directly add something in config/defaults.py ?

fmassa commented 5 years ago

By default, flipping is done automatically during training.

If you want to add more transformations, you can modify this file to add other transformations, like color transformations etc.

I at some point also added support for random crops, but because we need to handle specially the cases of no boxes, I removed it. But once we handle that better, I can add it as well.

txytju commented 5 years ago

@fmassa Can you handle no-boxes-cases now? I'm willing to work on that.

fmassa commented 5 years ago

@txytju I have not myself added the no-boxes handling, but an user has implemented it in https://github.com/facebookresearch/maskrcnn-benchmark/issues/169#issuecomment-447306468

Could give it a try and let me know?

Thanks