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.
The task is to localize the wheat head contained in each image.
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.
.
├── 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
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.
Use Filtering-Outputs for filtering outputs during inference
Use TTA for Test-time-augmentation
Pytorch-vision docs
Pytorch-vision github