First of all, I would like to thank Zindi, Wadhwani AI for this great competition.

My solution is quite simple, just need a good understanding of the object detection framework.

I use DINO from the detrex repository.

I trained 2 models, dino_swin_base_384_4scale_12ep and dino_swin_base_384_4scale_36ep with all data, including the negative samples, to make the model well generalized.

The visualization code was then tweaked to create the submission file, ensemble by WBF.

Since I only train the model with 384x384 resolution but the score is quite consistent across public and private datasets, I think improving the resolution can achieve better performance. Unfortunately, due to the limitations of the hardware, I cannot perform it.

Thanks for reading.