tlpss/ap10k-cropped-512x512

AP-10K, instance-cropped to 512x512

A repackaging of AP-10K for use in 2D-keypoints-benchmark. The images and annotations are those of AP-10K; only the packaging differs.

What was changed

The upstream dataset declares 54 categories (one per species) that all share the same 17-keypoint quadruped skeleton, and contains ~1.3 annotated animals per image. Two transformations were applied:

1. The 54 species categories were collapsed into a single animal category, keeping the shared 17 keypoints and skeleton. Species identity is not preserved in these annotations.
2. Every instance was cropped into its own image: the bounding box was expanded by 25%, made square, shifted to lie inside the source image where it fits and zero-padded otherwise, then resized to 512x512. Keypoints falling outside a crop are marked as not labeled (v=0).

The result is single-category and single-instance, in COCO keypoints format, with train/val/test taken from the official split1.

Note that cropping to ground-truth boxes makes this a top-down task that assumes a perfect detector. It is not suitable for evaluating detection or multi-instance association.

Splits

Derived from AP-10K split1 (7,023 / 995 / 1,997 source images; 9,122 / 1,272 / 2,634 instances). Instances left with no visible keypoints after cropping are dropped.

License and attribution

CC-BY-4.0, inherited from AP-10K. If you use this data, cite the original:

@inproceedings{yu2021ap10k,
  title     = {{AP-10K}: A Benchmark for Animal Pose Estimation in the Wild},
  author    = {Yu, Hang and Xu, Yufei and Zhang, Jing and Zhao, Wei and Guan, Ziyu and Tao, Dacheng},
  booktitle = {Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2021}
}