This demo is a proof of concept for scene-aware car compositing with specialized fine-tuned detectors and lightweight geometry logic.
Fine-tuned YOLOv11 model on ~4k labeled images for robust window-region detection/removal in interior scenes.
Fine-tuned YOLO26n model on ~1k labeled images for wheel-point detection in exterior scenes.
Wheel keypoints are used to infer coarse vehicle orientation and grounding, then adapt placement and contact-shadow behavior to the target background geometry.
This is currently suboptimal: using more anchor/key points on the car body would improve reorientation, scene placement consistency, and shadow accuracy.
Optional SAM3 detection can constrain/refine exterior cutouts when base segmentation is unstable.
Uses serverless SAM3 car detection for exterior cutout guard (configured via environment variables).
If scene is exterior: detects tire keypoints with `checkpoint/pose/best.pt`, centers/scales car in scene and
adds a light-direction-aware shadow from the selected background.
Runs several model passes and combines them to reduce tiny misses/flicker.
Turn off for pure single-pass output.