Training YOLOv5 w/ SAHI on TINY (< 5x5 pixels) object detection

[ansib1e]

Hello,

I'd firstly like to thank this community for this awesome work and in advance for your help. I'm trying to train a YOLOv5 model to use with SAHI for detecting tiny cars in satellite imagery, but the size of the cars are extremely small relative to the imagery (sometimes less than 5x5 pixels due to satellite resolution). The original annotated images are on average larger than 1280x1280 pixels so the annotations look like tiny specks and fill less than 1% of the full-sized training images. I'm wondering what the best approach would be for training/fine tuning?

I've trained a YOLOv5 model on similar, but higher resolution imagery in the past for another project with no issues and SAHI was able to detect all the cars in extremely large input imagery, but the annotated objects in that case were much larger relative to the slicing window (usually larger than 25x25pixels).

I'm currently training on 640x640pixel slices with a YOLOv5l model without much success and am wondering if I should try increasing the image sizes to 1280x1280 or try a different model with SAHI? I've seen in the published paper that this exact problem was addressed with slicing aided fine-tuning, which I believe involves using the SAHI slicing command to prepare sliced chips of the training imagery and am wondering if I am missing something here? I've tried slicing the training chips (640x640) into windows training on that dataset, but the objects are still too small relative to the sliced 640x640 pixel training chips. I understand this may just be a limitation of the resolution of my training imagery, just wanted to try posting here for some help from the community. Thanks again in advance for all the help!