Congratulations to all winners!!! This competition was amazing

Here is my solution (PUBLIC 0.330434782 PRIVATE 0.322933643)

Preprocessing:

- Delete images without annotations.

- Make a function that detects images whose pixels have been completed by the drone. These images stand out by a significant presence of pixels with an intensity = 0 or 255. I deleted the images containing 0.9 white or black pixels.

- 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.

- Put on the COCO format

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

Training:

- Build a CustomDataset and perform data augmentation (resize to 800*800, hlip, vflip, randombrightnesscontrast, colorjitter, normalization with value of IMAGENET). I use 800*800 because we have small object and for me to help the model a big resolution was good. The Batch size choose was 16.

- Focal Loss

- Faster RCNN with Mobilenet large FPN. I perform the model on the train dataset on all the classes. SGD with a lr = 5e-3 was used, momentum = 0.9, weight decay = 5e-4. 50 epochs for the iterations.

- I combine the dataset and here a fine tune with the pretrained params using a Adam optimizer with lr = 2e-4, I used 30 epochs.

Experience:

After fine tuning, I take a conf = 0.45 of the model to make the inference on the test data.

I also used DeTr but I haven't good result like Faster RCNN (PUBLIC 0.36). I think for this problem convolution was better than transformers.

Final Comment:

Imbalance is clearly the major challenge of this challenge as well as small houses. Unfortunately my work was not enough to manage the problem. Thank you all,

For those who would like or who could enlighten me to improve my work, I am completely open because it leaves me, I'm not going to lie to you, with a bitter taste because I don't understand my results.

Thank you and have a good day, see you in the next challenge