Discussion on "YOLO from Scratch: Object Detection"

DataPi ML · 2026-10-04T09:30:00.000Z

When we look at an image, we do not only recognize which objects appear in it; we are also able to locate where they are. For a computer vision model, solving both tasks at the same time is an object

The export section is where I would add a coordinate-level check. ONNX or TensorRT execution can produce convincing detections while the application maps boxes incorrectly after resizing and padding. The normalized and pixel-coordinate representations you explain earlier make a useful test fixture for that boundary.

Run a few non-square images through both the original and exported pipelines, then compare boxes in the original image coordinates with identical preprocessing and suppression settings. Include objects near the padded edges and predictions close to the confidence threshold. That catches integration mistakes that an aggregate speed measurement will miss, and separates export differences from postprocessing differences.

Thanks for the detailed suggestion. That’s a really useful point, especially the distinction between export issues and preprocessing/postprocessing mistakes. Comparing box coordinates on non-square images and around padded regions is a great validation step that can easily be overlooked. We’ll definitely keep it in mind for a future article focused on deployment and inference optimization.

That deployment article could make the boundary very concrete with one fixture that records the resize ratio, padding offsets and expected boxes in original coordinates. I would compare the outputs before suppression as well as the final detections, so a threshold or NMS difference does not hide a coordinate mismatch.

Keeping the fixture in the exported runtime’s integration checks would also make it useful after model updates, when the application mapping may remain unchanged but the output layout or precision changes.

Your 'detection is not tracking' section is where I'd want a metric for identity. I score agents step by step because one good aggregate can hide a broken stage, and per-frame mAP looks like that trap. Will the tracking article cover ID switches or IDF1?

Search Hashnode

YOLO from Scratch: Object Detection

Responses4