Unreal Engine 5 synthetic data for object detection: 1,124 bounding boxes, none drawn by hand
retained text
Every box in the clip comes out of the engine at render time. Each object gets its own ID in a separate buffer, and its box is measured from those pixels. Nobody drew any of them. 14 frames from 7 maps, 1,124 boxes.
What's in the video:
- The line sweeping across is only there for the video. As it passes, each object's mask flashes, then its box appears.
- Colour is the class: pink for people, and chairs, benches, umbrellas, wheelchairs, bikes, cars, tractors, hay bales and animals each get their own.
- A box covers what the camera can actually see. Someone half behind an umbrella gets a box around the visible half.
- The beach is the newest scene: 221 people placed per frame, standing, lying down, in deck chairs and wheelchairs, shot from 07:00 to 18:30.
It exports YOLO and COCO, with segmentation masks if you want them.
It's still rendered people though, and that costs a lot on real footage: the last detector I trained only on renders got 0.35 recall on real drone video. Closing that gap is what I'm working on now.
Which real dataset would you test this against? I'm most interested in crowded ones where people overlap, since that's the slowest part to label by hand.document fetched 2026-09-27 00:23 · canonical https://reddit.com/r/computervision/comments/1wqmra6/unreal_engine_5_synthetic_data_for_object