ChildPose

Model weights and inference code for ChildPose: Foundation for Children Pose Modeling (NeurIPS 2026). Code repository: jiaka1chen/ChildPose. Project page: pediamed.ai/ChildPose.

ChildPose estimates COCO-17 2D keypoints of a child in video. It is a top-down model: each frame is cropped around the child's bounding box, encoded by an image backbone, conditioned on a memory of the child's previous frames, and decoded into keypoint heatmaps. The memory only looks back a fixed number of frames, so videos of any length are processed frame by frame.

Files

Variant	Backbone	Input (H x W)	Parameters	Checkpoint
hrnet_w48	HRNet-W48	288 x 384	102.5M	checkpoints/childpose_hrnet_w48.safetensors (0.41 GB)
hrformer_base	HRFormer-B	288 x 384	82.1M	checkpoints/childpose_hrformer_base.safetensors (0.33 GB)
vitpose_base	ViTPose-B	256 x 192	108.2M	checkpoints/childpose_vitpose_base.safetensors (0.43 GB)
sapiens_0.3b	Sapiens-0.3B	1024 x 1024	346.1M	checkpoints/childpose_sapiens_0.3b.safetensors (1.38 GB)
sapiens_0.6b	Sapiens-0.6B	1024 x 1024	696.3M	checkpoints/childpose_sapiens_0.6b.safetensors (2.79 GB)
sapiens_1b	Sapiens-1B	1024 x 1024	1227.3M	checkpoints/childpose_sapiens_1b.safetensors (4.91 GB)

Each checkpoint is a safetensors file with float32 weights; the model configuration is stored in its metadata. The inference code (childpose/, scripts/) is included in this repository and is the same as in the code repository.

Usage

git clone https://huggingface.co/Asp1rinn/ChildPose   # needs git-lfs; or download single files
pip install -r requirements.txt

from childpose import ChildPosePredictor

predictor = ChildPosePredictor("checkpoints/childpose_vitpose_base.safetensors")
# or ChildPosePredictor("vitpose_base"): downloads only that checkpoint from this repository

# frames: any iterable of PIL images or RGB uint8 arrays (a generator is fine for long videos)
# boxes:  per-frame child box [x1, y1, x2, y2] in pixels, as a list or {frame_index: box}; None = not visible
results = predictor.predict_video(frames, boxes, fps=30)
# results[i]: None, or {"keypoints": (17, 2) image pixels, "scores": (17,), "score": float, "crop_box": [...]}

Command line, for a video and a JSON file of per-frame boxes for each tracked child:

python scripts/run_video.py --checkpoint checkpoints/childpose_vitpose_base.safetensors \
    --video clip.mp4 --boxes boxes.json --out poses.json --vis poses.mp4

{"child_0": {"0": [100, 50, 300, 450], "1": [102, 51, 301, 452]},
 "child_1": {"0": [500, 80, 640, 400]}}

Input

• Boxes: tight [x1, y1, x2, y2] boxes around the child in each frame; the model enlarges them by 10% before cropping. ChildPose does not detect or track children; use any person detector / tracker.
• Frame rate: the models were trained on frames sampled at about 5 fps; each frame attends to up to 6 earlier frames, about 0.2 s apart in training. The video is processed in one pass in frame order and the memory runs over the whole video. For fps=F, mode sets how frames reach the memory, with step = round(F / 5): "interleave" (default) predicts every frame, frame i remembering frames i - step, i - 2*step, ... (about 0.2 s, 0.4 s, ... earlier); "downsample" predicts only every step-th frame (the video at about 5 fps); "sequential" predicts every frame with a memory of the immediately preceding frames. Leave fps=None for input that is already at about 5 fps.
• use_memory=False predicts every frame on its own; predict_images handles single images.

Output

COCO-17 keypoints in image pixels (order in childpose.COCO_KEYPOINTS), with a per-keypoint score (the heatmap peak value).

License

The weights and code are released under CC BY-NC 4.0 (see LICENSE), the most restrictive license among the components they build on (Meta's Sapiens is CC BY-NC 4.0). Third-party code and weight notices (ViTPose, MMPretrain: Apache-2.0; HRNet, HRFormer, openseg.pytorch: MIT) are in THIRD_PARTY_NOTICES.md.

Responsible use

Use of the code and weights is subject to the ChildPose Responsible Use Policy. In short:

• Intended use: non-commercial research on children's pose and movement, on lawfully recorded video with the consent of the children's parents or legal guardians, under human oversight.
• Prohibited: re-identification or identity linkage (including from body shape or gait); covert surveillance or monitoring of children; use on its own for clinical diagnosis, screening, treatment or any other high-stakes decision about a child; commercial use; any use that violates the law.
• Not a medical device: not clinically validated or approved by any regulatory authority. Predictions can be wrong, and performance across skin tone, body habitus, clinical condition and capture setting has not been validated. Deployment in medical, educational or monitoring settings needs its own validation, fairness evaluation, privacy safeguards and human oversight.
• No warranty, no liability: the models are provided "as is"; the authors and their institutions are not liable for any use of them or of their outputs (see LICENSE and the policy).

Citation

@inproceedings{chen2026childpose,
  title     = {ChildPose: Foundation for Children Pose Modeling},
  author    = {Chen, Jiakai and Shen, Yifan and Li, Boyi and Dong, Chuanmiao and Cao, Xu},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}