Assign a purpose to each batch before splitting samples. Synthetic images can supplement training and support screening or stress tests under controlled conditions. Real validation samples can guide parameter selection. Reserve a separate real test set for final evaluation. For sites with limited samples, record the cameras, operating conditions and recording periods covered, along with any missing conditions.
Split samples by source group
Keep related samples in the same source group when the evaluation objective calls for it. Examples include adjacent frames from one video, edits of one original image and variants generated from one base image. To assess new sites, consider grouping by site. In grouped cross-validation, each source group belongs to only one of the two sets within a split. The two sets are for training and validation. See the grouped cross-validation documentation.
Save the split manifest
Record sample identifiers, source groups, real or synthetic origin, assigned purpose and label version. Check for overlapping groups and near-duplicate images after splitting. Complete model selection and threshold tuning before using the reserved real test set. Interpret results within the sample scope and conditions. Record synthetic stress tests separately.
Related services: Data, training and evaluation, project consultation.