Understand image data
Inspect width, height, channels and numeric range before processing images. Convert colour spaces deliberately and keep resizing consistent between training and inference. Display a sample after each preprocessing step; a channel-order error can produce a technically valid array with incorrect meaning.
Choose the task
Classification assigns an image label, detection locates objects with boxes, and segmentation labels pixels. Choose annotations that fit the task. Do not evaluate a detector as though it were a classifier: localization quality and missed objects matter as well as category predictions.
Prevent data leakage
Keep images from the same source, video or person in one split when they are strongly related. Randomly splitting adjacent video frames can inflate results. Check class balance and evaluate difficult lighting, occlusion and unfamiliar backgrounds separately.
Practice project
Build a small object-counting demonstration using data you are allowed to use. Start with a manually checked validation set. Report precision, recall and examples of missed or duplicated objects. Deliver annotation notes, inference instructions and a failure gallery rather than an unsupported deployment claim.
Computer vision lab: evaluate difficult examples
Make a separate evaluation folder containing low light, partially hidden and unusually oriented objects. Keep the original images and document transformations. Run the same model and threshold over every category. Count missed objects and false detections manually for this small set. Explain why a threshold that improves recall may also increase false positives. A useful report includes several annotated mistakes and a clear statement of the conditions that were not tested.
Check your progress
Keep your working files, test cases and notes together. Explain one design decision, demonstrate a failure case and list what you would improve next. These concrete outputs make it easier to identify the next skill to practice.