Choose the output and annotation
Image classification assigns one or more labels to an image. Object detection identifies objects and their locations, often with bounding boxes. Define whether multiple objects or classes can occur, what counts as a positive example, and how uncertain or partially visible objects should be labeled. Review annotation instructions and disagreements before training; a model cannot repair an inconsistent target definition.
A factory needs to know whether a package image contains damage and where the damage is located. Explain why an image-level label alone may not satisfy the location requirement, and describe one annotation edge case.
Split related images together
Random image-level splits can put near-duplicate frames, images of the same item, or repeated capture sessions in both training and test sets. The resulting score may reflect memorized backgrounds or objects instead of performance on genuinely new cases. Group by the relevant unit, such as item, patient, camera session, or source video, and consider a later-time or new-location holdout when that matches deployment. Keep augmented versions of an original image in the same split.
A dataset has 100 video clips with 50 frames each. Explain why randomly splitting all 5,000 frames can leak information and propose a clip-level split. What additional test would assess a new camera location?
Evaluate the conditions that matter
Inspect performance by class and relevant capture conditions such as lighting, camera, resolution, occlusion, and location. For classification, examine confusion matrices and precision and recall at a selected threshold. Detection also requires evaluating localization and missed or duplicate objects using suitable task metrics. Report counts and uncertainty for small slices. Compare with the operational costs of a false alarm or a missed defect, not just a single aggregate score.
A model performs well on daylight images but misses defects at night. Identify a possible data-coverage issue, a slice-level metric to report, and a safe response before expanding deployment.
Test the production image path
Document image orientation, resizing, color handling, compression, and any transformations used by the model and serving application. Compare representative images through training and serving paths, including boundary cases. Agent Platform supports managed image datasets and configurable splits for supported AutoML workflows; check current requirements, regions, costs, and annotation formats in the official documentation. Monitor capture changes and review errors after deployment.
An upstream service rotates phone images based on metadata, but the training pipeline did not. Design a parity test with a rotated sample and specify what output or intermediate representation you would compare.
Many frames from the same video appear in both training and test sets. What is the main concern?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.