EuTECSSG / cvpr_wad

This is the repository including our contribution to cvpr_2018_autonomous_driving_challenge, forked from our team's repository.

MASK_rcnn_train_custom_data

Installation

1. clone this repo:

git clone [e-mail]:kagglewadteam/cvpr_wad.git --recursive

2. see https://github.com/matterport/Mask_RCNN#installation

Directory Structure

.
├─assets
├─dist
├─images
├─mask_rcnn.egg-info
├─mrcnn
├─mask_rcnn_coco.h5 ❤️
└─samples
  └── .ipynb_checkpoints
  └── balloon
  └── coco
  └── kaggle_wad
  └── nucleus
  └── shapes

the files with heart are newly added. Not neccessary to download the mask_rcnn_coco.h5
it can be downloaded automatically.

Training

cd into the kaggle_wad directory

cd Mask_RCNN/samples/kaggle_wad
python3 train.py

one thing should be noticed that your own directories of images and masks should be defined in the train.py

Tuning hyper-parameters

└─mrcnn
  └── config.py ❤️
  └── model.py
  └── parallel_model.py
  └── utils.py
  └── visualize.py
  └── init.py

most parameters in the config.py, few in the kaggle_wad/train.py

Detection

cd into the kaggle_wad directory

cd Mask_RCNN/samples/kaggle_wad
python3 detect.py

Evaluate Submission

Validate the format of submission csv file.

1. cd into the submission directory:

cd submission

2. Change the directory of test images and submission file accordingly
3. Validate the submission file

python3 submission_checker.py