Answers are grounded in this paper and cite their sources.
Synthetic object insertion is usually evaluated under one fine-tuning regime, leaving open whether its benefit depends on which detector parameters are updated. We compared 150 Unity-rendered vessels inserted into xView training images with an exposure-matched Duplicate control under three adaptation scopes (Detect-only, Neck+Detect, or full network) over ten paired seeds from xView-adapted YOLO26 checkpoints. In YOLO26n on xView, within one execution environment, the Unity-minus-Duplicate vessel average precision (AP50–95) fell from +0.71 percentage points (pp; Detect-only) to +0.10 pp (Neck+Detect) and −0.99 pp (full network), a Detect-minus-Full interaction of +1.70 pp (95% confidence interval [+1.32, +2.08]). Quality-controlled real insertions showed a smaller interaction; no differential forgetting was detected. The Detect-only benefit and full-network loss were replicated on xView validation scenes, whereas the interaction was not reproduced when starting from COCO-pretrained weights (20 or 50 epochs), indicating sensitivity to initialization and training schedule. Interactions were inconclusive for YOLO26s, YOLO26m, and YOLO11n, and xView has four vessel-positive test scenes; on 343 prospectively locked DOTA images, all YOLO26 interactions were positive, but Neck+Detect gained most, so scope ordering is domain-dependent. These are controlled contrasts at low absolute AP; synthetic augmentation should be reported with adaptation scope, initialization, schedule, evaluation domain, and execution environment.
No takes yet. Share an insight, caveat, or question.
Choi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: