mapping-challenge-starter-kit

Installation

git clone https://github.com/crowdAI/mapping-challenge-starter-kit
cd mapping-challenge-starter-kit
pip install -r requirements.txt

Dataset

Please download the datasets from https://www.crowdai.org/challenges/mapping-challenge/dataset_files, and put them in the data/ folder. Untar them (this might take some time) to have the following directory structure:

|-- data/
|   |-- test_images/ (has all images for prediction)
|   |-- train/
|   |   |-- images (has all the images for training)
|   |   |__ annotation.json : Annotation of the data in MS COCO format
|   |   |__ annotation-small.json : Smaller version of the "annotation.json"
|   |-- val/
|   |   |-- images (has all the images for training)
|   |   |__ annotation.json : Annotation of the data in MS COCO format
|   |   |__ annotation-small.json : Smaller version of the "annotation.json"

Usage

Now you can refer to the list of Jupyter Notebooks for different aspects of the challenge and the datasets. You can access all of them by :

jupyter-notebook

Available Notebooks

• Dataset Utils

  • Import Dependencies
  • Configuration Variables
  • Parsing Annotations
  • Collecting and Visualizing Images
  • Understanding Annotations
  • Visualizing Annotations
  • Advanced
    • Convert poly segmentation to rle
    • Convert segmentation to pixel level masks

• Random Submission

  • Submission Format
  • Generating a Random Segmentation
  • Generating a Random Annotation Object
  • Generating a Random Results Object
  • Submit to crowdAI for grading

• Locally test the evaluation function

• Train Mask-RCNN

  • Installation
  • Training
  • Prediction & Submission
  • NOTE : This is in a separate repository, and we have also now added the pretrained weights from the baseline submission to the datasets page.

Miscelaneous Resources

• Convert Annotations from MS COCO format to PascalVOC format

Acknowledgements

A big shout out to our awesome community members @MasterScat (Florian Laurent), Snigdha Dagar, and Iuliana Voinea, for their help in preparing the datasets and designing the challenge.

Author

Sharada Mohanty [email protected]