DensePose-COCO: Human Image-to-Surface Correspondence Annotations

DensePose-COCO is a large-scale ground-truth dataset with image-to-surface correspondences manually annotated on COCO images. It contains 33,929 samples and was created by Voxel51. The dataset was last updated on the Hugging Face platform in June 2024.

Use Cases

• Training dense human pose estimation models based on the manually annotated image-to-surface correspondences.
• Benchmarking 3D human body reconstruction algorithms based on the provided ground-truth correspondences.
• Developing models for human-centric tasks like action recognition or virtual try-on based on detailed body surface annotations.

Strengths

• 33,929 annotated samples provide a substantial base for model training.
• Manual annotation of image-to-surface correspondences offers precise ground truth.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Voxel51

Collection Method

Manually annotated image-to-surface correspondences on COCO images.

Freshness

Last updated 2024-06-20 17:16:33; freshness should be verified.