A Computer Vision Project Checklist

From pixels to production

Is the task the simplest one that answers the question (classification versus detection versus segmentation)? Do the images cover real conditions and devices, with clear labelling guidelines and leak-free splits? Is preprocessing (colour order, normalisation, resizing) identical in training and serving? Are metrics appropriate (accuracy, mAP, IoU or Dice) and reported per class, condition and relevant group? Is robustness tested? Does the model meet latency on the target hardware? Are privacy and legal questions settled, and is monitoring in place?

The checklist

Use it in design and launch reviews.

[ ] task chosen by the output needed (class / boxes / masks)
[ ] data covers real devices, lighting, angles; split by source
[ ] labelling guidelines + agreement measured
[ ] pretrained backbone + linear probe baseline before fine-tuning
[ ] preprocessing identical in training and serving (RGB, normalisation)
[ ] metrics per class / condition / group (accuracy, mAP, IoU, Dice)
[ ] robustness set: blur, noise, low light, new cameras
[ ] latency measured on target hardware
[ ] privacy: minimise, blur, retention, consent, legal check
[ ] monitoring per camera/site + golden-image tests

Start with a baseline

A pretrained model plus a linear probe often sets a high bar; fine-tune only if the baseline misses the target.

त्वरित जाँच: Which item belongs on a computer vision launch checklist?

• Identical preprocessing in training and serving, verified with golden images
• Feeding BGR images to an RGB model
• Splitting video frames randomly across train and test
• Skipping latency tests