The number of keypoints varies by dataset. For example, the LSP dataset has 14 keypoints, the MPII dataset has 16 keypoints, and the COCO dataset has 17 keypoints
COCO Dataset
To train on coco (single person pose estimation), place the following files in the right place.
cocoapi
: coco APIannotations
: COCO keypoint datasetperson_keypoints_train+val5k2014
: annotations (JSON files)
images
: tain and val datasetstrain2014
val2014
coco_annotations_2014.json
orcoco_annotations_2014.json
: reformatted annotation generated by./miscs/gen_coco.m
Prepare json
annotation file
Please edit year = 2017; % 2014 or 2017
in gen_coco.m to generate the corresponding annotation file (coco2014 or coco2017).
Dataset | Year | Labeled | Images | Keypoints | |||||||
---|---|---|---|---|---|---|---|---|---|---|---|
Humans | Train | Validation | Test | Body | Face | Hand | BBox | Mask | Activity | ||
COCO | 2014 | 250k | 64k | 17 | - | - | ✔ | ✔ | - | ||
MPII | 2014 | 40k | 25k | 7k | 16 | - | - | ✔ | - | ✔ | |
AIC | 2017 | 300k | 210k | 30k | 60k | 14 | - | - | ✔ | ||
MHP | 2018 | 15k | 5k | 5k | 16 | - | - | ✔ | |||
CrowdPose | 2018 | 10k | 2k | 8k | 14 | - | - | ✔ | |||
OC Human | 2019 | 62k | 2.5k | 2.2k | 17 | - | - | ✔ | |||
MPII-TRB | 2019 | 14 | - | - | ✔ | ||||||
Human-Art | 2023 | 123k | 35k | 5k | 10k | 17 | - | - | ✔ | ||
COCO-WholeBody | 2020 | 50k | 2k | - | 23 | 68 | 42 | ✔ | |||
Halpe | 2022 | 118k | 5k | 41k | 26 | 68 | 42 | ✔ |
Types of Keypoints and Skeleton Variants
How Different Datasets Define Keypoints and Joint Connections • Common skeleton structures in COCO, MPII, and Human3.6M • Differences in full-body vs. hand vs. face keypoints • Unified keypoint formats for multi-task models