Before adding images for underrepresented conditions, record the task, available real images and missing scenes. List the combinations of cameras, views, lighting, occlusion and object appearance that need more samples. Confirm permitted use of base images, assets and annotations before use. This is a proposed workflow. Its field performance is unverified.
Generate and screen images in batches
For each synthetic batch, retain base image sources, generation settings, scenes, a file list and label versions. First check that files match the metadata. Then inspect a subset of the images for edge artifacts, object positions, occlusion and label consistency. Record reasons for uncertain cases and exclude or correct them as needed. Preserve links between originals and variants for splitting and review.
Evaluate on real scene samples
Reserve a real test set with sources separate from training. Compare the baseline model and the model trained with added samples. Use consistent metrics, annotation rules and test inputs. Record changes in training settings. The evaluation report should include common and rare scenes, error examples and conditions with insufficient samples. If independent real images are unavailable, deliver the image generation and screening records first. Mark field performance as unverified.
Related services: Data, training and evaluation, project consultation.