Global-Wheat-Detection-2021

This project includes a solution to Global-Wheat-Challange-2021.
Also, It's a good source to get started with implementation of Object-Detection in PyTorch.

Task

The task is to localize the wheat head contained in each image.

Dataset

The training dataset will be the images acquired in Europe and Canada, which cover approximately 4000 images and the test dataset will be composed of the images from North America (except Canada), Asia, Oceania and Africa and covers approximately 2000 images. You can download the dataset from here.

Directory Structure

.
├── Dockerfile 
├── EDA.ipynb                    (Exploratory Data Analysis)
├── FRCNN_Resnet_inference.py    (inference)
├── FRCNN_Resnet_training.py     (training)
├── README.md
├── datasets.py 
├── detection                    (Helper functions for training)
├── models.py 
├── my_utils.py                  (Helper functions)
├── submissions
│   └── submission.csv           (sample-submission)
├── test                         (Download test-images and save them in this folder)
└── train
    ├── train                    (Download train-images and save them in this folder)
    └── train.csv

Evaluation Method

The metric used is the Average Domain Accuracy (ADA). Two boxes are matched if their Intersection over Union (IoU) is higher than a threshold of 0.5. Final Score by this approach is 56.2% ADA. Highest score in the competition was 70% ADA.

References

Use Filtering-Outputs for filtering outputs during inference
Use TTA for Test-time-augmentation
Pytorch-vision docs
Pytorch-vision github