Learn From
Human Action to Robot Learning
CURDIE3 dataset (opens in a new tab)
Example 1 of 3 — House interior
Image descriptions and full-size examples
Each six-panel comparison shows the rendered input with actor bounding boxes, COCO-17 body pose keypoints, depth in meters, separate actor segmentation masks, hand keypoints, and scene/camera metadata. Matching colors distinguish tracked people across the pose and mask panels.
1. House interior
Three people appear at different distances inside a house: a person bending forward in the foreground, one walking away in the middle, and one partly visible at the back. The pose and mask panels distinguish all three, while the depth panel separates the nearer subject from the room and farther people. The metadata identifies a house scene with three actors, frame 94, a free camera, and a focal length of 1616.2 pixels.
2. Distant subject
One person is visible far from the camera through successive doorways in a house. The small body skeleton and actor mask isolate that person; the depth panel also distinguishes the nearer doorframes from the distant subject. Hand landmarks are shown on the same person. The metadata identifies one actor, frame 2, an orbit camera, and a focal length of 1885.5 pixels.
View full-size distant-subject example
3. Outdoor field
Two people are in a grassy field. One leans back with bent legs in the foreground, while the other stands farther away. The foreground person is tracked in blue in the pose and mask panels; the standing person is tracked in green. The depth panel separates their distances, and hand landmarks mark both people's hands. The metadata identifies a field scene with two actors, frame 1976, a free camera, and a focal length of 3600 pixels.
A large-scale synthetic pose estimation dataset with 1.4M+ photorealistic images rendered in Unreal Engine. Features 100 actor models across 1,400 movements from real humans with pixel-perfect ground truth including COCO-17 body keypoints, MediaPipe hand keypoints, metric depth maps, and instance segmentation masks.
More datasets coming soon!
Contact Us