Synthetic data projects need delivery criteria for images and evaluation criteria for models. Image checks can cover files, dimensions, scene distribution, metadata, labels and exceptions. Model evaluation requires training settings, real samples and a defined test method. Specify whether delivery includes images, task annotations, or model training and evaluation.
Check the generation batch and missing items
The delivery checklist should record the planned number of images for each scene, actual files, generation versions, sources and screening results. Checks for occlusion, brightness or compression need defined samples and criteria. If delivery includes bounding boxes, segmentation masks or character labels, inspect examples against the task specification. Hold images for review if their origin or labels are missing.
Define a controlled model comparison
Use the agreed real test set to evaluate both models. Compare the original baseline with the model trained with added synthetic images. Use the same sample scope, annotation rules and evaluation method for both models. Record any differences in training settings. Examine false detections and missed detections in important scenes. Model benefits require comparative results. Untested field conditions remain pending verification.
Related services: Data, training and evaluation, project consultation.