Wheat Detection

The competition was hosted on Kaggle. The task was to detect and localize wheat heads in images.

Team Name : The Alphas

Authors : Ashay Ajbani and Pritam Rao

Competition Type : Object Detection

Framework : PyTorch

Solution Type : Transfer Learning

Number of Classes : 1

Pretrained Models Used : 
1) Detection Transformer (Submission 1)
2) Efficient Detector (Submission 2)

Public Leaderboard Rank : 481/2270

Private Leaderboard Rank : 354/2270

Our Approach

Submission 1 :

For our first submission, we fine tuned pretrained Efficient Detector model for our problem. It gave us a score of 0.7331 on the public leaderboard.

Click here to access the code

The file wheat_detection_efficientdet.ipynb includes training as well as inference. We have used Test Time Augmentation and Weighted Boxes Fusion for making more accurate predictions.

Submission 2 :

For our second submission, we fine tuned pretrained DETR model for our problem. It gave us a score of 0.5758 on the public leaderboard.

Click here to access the code

We have uploaded the following DETR notebooks :

1) wheat_detection_detr_training.ipynb : Includes processing the dataset and training on DETR.

2) wheat_detection_detr_inference.ipynb : Making predictions on the test images using the trained model.

We have documented the notebooks so that you can easily reproduce the results.