Fine-tuning a COCO-pretrained model on 1,600 Vivid 3D synthetic renders improved detection AP50:95 by 4.45 percentage points and mean bounding-box overlap by 1.52 percentage points in a targeted real-photo test.
Test scope: 34 real photos, 43 annotated cars, the same test for both models.
Read the result correctly: this is a targeted car-detection benchmark, not a claim of better detection across all cars or environments.
Case type: internal study run by Vivid 3D. Industry: automotive. Use case: object detection. Pipeline used: 3D asset creation, asset management, synthetic data generation, model training and validation.
A COCO-pretrained detector already had general knowledge of cars. This experiment tested whether synthetic images from an existing 3D asset could improve the localization of cars in real photographs without adding real images during the additional fine-tuning stage.
The research question: can synthetic data improve vehicle localization in real photographs? The focus was box placement, not vehicle-make recognition, model classification or higher detection recall.
Each image was paired with a corresponding object annotation in Vivid 3D's synthetic training dataset. Scenes and camera viewpoints vary across the training set.
The comparison kept the real-photo evaluation set fixed and measured changes after synthetic-only fine-tuning of the pretrained detector.
Important distinction: no real photos were added in the fine-tuning stage. The starting model was already COCO-pretrained, so this was not training a detector from scratch without real data.
Both models detected the same 42 of 43 annotated cars. The measured gains came from better localization and detection quality across overlap thresholds, not from finding more cars.
pp = percentage points. IoU is intersection over union (100% indicates a perfect match). False detections are counted at confidence threshold 0.50.
In this single illustrative example (Photo 5), the predicted bounding box moved closer to the reference annotation. The crop and scale are identical in both panels.
COCO car AP50:95: 49.91% → 46.49% (−3.42 percentage points).
The broader COCO car benchmark became worse after fine-tuning. These findings support task-specific adaptation, not general improvement across cars, cameras, environments or deployment conditions.
This small test provides descriptive evidence only. Statistical significance and production-wide performance were not established.
It depends on the task and how far it is from what the model already knows. In this test, 1,600 renders of one 3D asset, plus 400 for validation, improved AP50:95 on real photos by 4.45 pp.
For a targeted task, yes. Here a detector fine-tuned only on synthetic renders improved on real car photos, while general COCO car detection declined by 3.42 pp.
Not in the additional fine-tuning stage. The original detector had already been pretrained on COCO.
No. The targeted real-photo benchmark improved while broader COCO car AP50:95 declined.
Real-photo test: 34 reviewed photographs with 43 annotated cars; Photo 41 excluded from both runs because of an incorrect annotation. The same set was used in earlier comparisons. Selected checkpoint: synthetic-only SGD, epoch 26. Reported gains are descriptive, not a statistical-significance claim.
Real-photo dataset: car-image dataset by tink, via Roboflow Universe (CC BY 4.0). Source photo cropped and annotated for comparison. This independent benchmark does not establish involvement or endorsement by any vehicle manufacturer.
Related reading: Synthetic Data in Computer Vision Annotation and Synthetic Data Generation Methods. More projects are in Vivid 3D case studies.
Align target classes, assets and real-photo validation criteria with the Vivid 3D team. Discuss your dataset.
