Synthetic Human Pose Images with Depth Maps and Keypoints

CURDIE3 is a synthetic computer vision dataset containing 1,408,410 photorealistic RGB images with multi-format ground truth labels. It features 100 distinct actor models rendered across three environments in Unreal Engine at 30 frames per second.

Use Cases

• Train a multi-task model on RGB images to jointly predict COCO keypoints, metric depth maps, and instance segmentation masks.
• Develop and benchmark depth estimation algorithms using the provided metric depth maps as ground truth.
• Use the synthetic instance segmentation masks for training or evaluating segmentation models without real-world data collection.
• Leverage the 100 distinct actor models and 1.4 million images for data augmentation or pre-training pose estimation networks.

Strengths

• Contains 1,408,410 images, providing a large-scale resource for training data-intensive models.
• Includes multi-format annotations (RGB, depth, segmentation, keypoints) for each image, enabling diverse computer vision tasks.
• Features 100 distinct human actor models, increasing the diversity of represented individuals.
• Rendered at 30 fps in Unreal Engine, ensuring high visual fidelity and consistent synthetic ground truth.

Limitations

• Data is entirely synthetic, which may introduce a domain gap when applying models to real-world imagery.
• The specific environments and actor demographics used may not represent the full diversity of real-world scenes and people.

Provenance

Source

Hugging Face user kikitora.

Collection Method

Photorealistic images and ground truth labels synthetically rendered using Unreal Engine.

Time Range

null

Freshness

Last updated on 2026-01 31.

Geography

Region tag indicates 'US', but the data is synthetic.