Datasets:

LibreYOLO
/

ap10k-pose

license: cc-by-4.0
task_categories:
  - keypoint-detection
tags:
  - libreyolo
  - pose
  - animal-pose
  - ap10k
pretty_name: AP-10K Animal Pose (LibreYOLO)
size_categories:
  - 1K<n<10K

AP-10K Animal Pose (LibreYOLO)

Multi-class animal pose in LibreYOLO YOLO-pose layout: detect and classify each animal into one of 54 species while predicting a single shared 17-keypoint quadruped skeleton. Hosted so model.train(...) / model.val(...) can consume it directly for multi-class pose (one kpt_shape for every class).

Provenance

Source: AP-10K, A Benchmark for Animal Pose Estimation in the Wild (Yu et al., NeurIPS 2021 Datasets and Benchmarks Track). Upstream: https://github.com/AlexTheBad/AP-10K (canonical release, Google Drive labeled set).
Pinned source sha256: 420980abb135d6f66bcc8e29f289a46081214016192ae197ad24bc1525c8e62c (the upstream ap-10k archive).
Transform applied: extracted the archive; converted the COCO-style keypoint split (split1) to YOLO-pose TXT, writing one class per species over the shared 17-keypoint skeleton; copied images into images/{train,val}; generated ap10k-pose.yaml (kpt_shape, flip_idx, skeleton, oks_sigmas). No third-party repackaging used.

ap10k-pose/
├── images/train/*.jpg   (7023)
├── images/val/*.jpg     (995)
├── labels/train/*.txt   (7023)   # <cls> <cx> <cy> <w> <h> (<x> <y> <v>)*17, normalized
├── labels/val/*.txt     (995)
└── ap10k-pose.yaml

Splits (AP-10K split1): 7,023 train images / 9,122 instances, 995 val images / 1,272 instances.
Classes: 54 species across 23 families (Bovidae, Canidae, Castoridae, Cercopithecidae, Cervidae, Cricetidae, Elephantidae, Equidae, Felidae, Giraffidae, Hippopotamidae, Hominidae, Leporidae, Mephitidae, Muridae, Mustelidae, Procyonidae, Rhinocerotidae, Sciuridae, Suidae, Talpidae, Ursidae, Vespertilionidae). Class index = species, ordered by upstream category id (see names in the yaml).
Keypoints (17, shared): left_eye, right_eye, nose, neck, root_of_tail, left/right shoulder, elbow, front_paw, left/right hip, knee, back_paw.

Use with LibreYOLO

from libreyolo import LibreYOLONAS
model = LibreYOLONAS("yolo_nas_pose_s_coco_pose.pth", size="s", task="pose")
model.train(data="ap10k-pose.yaml", epochs=100, imgsz=640)
model.val(data="ap10k-pose.yaml")

License

Source license: CC BY 4.0 (Creative Commons Attribution 4.0 International), inherited. Please attribute AP-10K:

Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, Dacheng Tao. "AP-10K: A Benchmark for Animal Pose Estimation in the Wild." NeurIPS 2021 Datasets and Benchmarks Track. https://github.com/AlexTheBad/AP-10K

AP-10K Animal Pose (LibreYOLO)

Provenance

Contents

Use with LibreYOLO

License