Developer fixes livestock counting by adjusting image processing, not AI model

A developer building an on-device livestock counting app encountered inaccurate results when testing on a real photo containing roughly 250 sheep. The CoreML model counted only 14 animals because the high-resolution image was being resized too small for the detector to recognize distant sheep. The solution involved processing the image in overlapping tiles so the model could see objects at a larger scale, without altering the model itself. This tiling technique increased detections from 14 to 173 on the same photo while maintaining accuracy on close-up images. The experience underscored that image pipeline adjustments often yield significant improvements before considering model changes.