Preprocessing:

- Delete images without annotations

- Delete mal annoted images (3 images). bbox is [x,y,w,h], if the area of the bbox which is a rectangle is null so w * h = 0 we have problem.

- Split the dataset (80% for the train and 20% for the validation)

Training:

- Used a Faster RCNN model with a resnet152 backbone. Treated the problem as an object detection task

-Trained for 2 epochs, batch size of 4, did not resize the images, SGD optimizer, 0.001 lr

- Model trained for about 5 hours

Post-Processing:

Post processing is very key to this competetion. I tried several confidence scores but I finally settled with 0.83 threshold.

Final Comment:

Faster RCNN models did better than the YOLO models. Post-processing is very vital to a successful model.

Thank you