Detectron: How can I make my own dataset for mask RCNN training ?

[how0723]

How can I make my own dataset for mask RCNN training ?

[chrisby]

From my understanding of the infrastructure of the framework I would say there are 4 essential steps to train a mask RCNN with detectron:

1. Create an annotation file for your data set and set up symlinks as described in [data README].(https://github.com/facebookresearch/Detectron/blob/master/lib/datasets/data/README.md) Ideally you use the coco API.
2. Choose the baseline architecture from the model zoo, adjust respective config file to your needs (e.g. /configs/12_2017_baselines/mask_rcnn_R-101-FPN_1x.yaml) and train the model with tools/train_net.py [OPTIONS]
3. Train an RPN (e.g. /configs/12_2017_baselines/rpn_R-101-FPN_1x.yaml) to receive a model that then creates (by inference) a proposal file on your test set.
4. Use created proposal file for inference of the model trained in 2).

I did not go through steps 3 and 4 yet, but I think this would be the workflow.

[groupbox-dev]

This might provide relevant solutions to the underlying problemhttps://github.com/ppwwyyxx/tensorpack/issues/632

[waspinator]

I wrote a library and article to help with creating COCO style datasets.

https://patrickwasp.com/create-your-own-coco-style-dataset/

[pascschoSSL]

@chrisby how did you do step 4? I trained my mask_rcnn and also an RPN on my own dataset. Then created (by Inference) the detection.pkl file (proposal file) (like step 1-3 in your description). But now I am not sure how to include the proposal file (of my test set) for inference of the model trained in step 2.

[chrisby]

@chrisby how did you do step 4? I trained my mask_rcnn and also an RPN on my own dataset. Then created (by Inference) the detection.pkl file (proposal file) (like step 1-3 in your description). But now I am not sure how to include the proposal file (of my test set) for inference of the model trained in step 2.

Unfortunately, I had to drop my project but I assume you put the generated proposal file into the TEST -> PROPOSAL_FILES part of the .yaml (e.g. line 40 of https://github.com/facebookresearch/Detectron/blob/master/configs/12_2017_baselines/mask_rcnn_R-101-FPN_2x.yaml)