Hey,
Not sure if this is he right place to ask, but giving it a shot. Looking for some help, tips or tricks on building a robust and accurate model for pellet hole detection to analyze swarm patterns.
I have so far relied on open cv and got decent results. However I do need to improve accuracy and especially recall, so I have started labelling images and started training a yolo nano model. the model has surpassed the open cv image processing algorithm, but I would prefer even stronger results.
I am not very up to speed with the state of the art methods and was wondering if you have any thoughts on what methods I should try out. Bigger model, more data, etc..?
Happy to share more details around my current set up. Here are some photos to give you a feel for my problem.
Yes — this is a very good Ultralytics YOLO use case, and pellet holes are basically a small-object detection problem.
The biggest gains usually come from making the holes larger relative to the input, not just using a bigger model. I’d try Ultralytics YOLO26 with yolo26s.pt or yolo26m.pt, train at a higher imgsz like 1280 or 1600, and preferably split each full sheet into overlapping tiles before training/inference so each hole occupies more pixels. The guidance in this small-object detection article for YOLO11 still applies well here: use light augmentation, avoid transforms that erase tiny objects, and review misses carefully.
If some holes are ragged or overlapping, segmentation may work even better than boxes for counting. I’d also make sure labels are extremely consistent, since tiny annotation errors hurt recall a lot.
If you want, share your current model, imgsz, dataset size, average hole size in pixels, and a couple of failure cases, and I can suggest a tighter setup.
