baki0115/GazeCtrl_dataset

GazeCtrl — human-curated gaze labels for body reorientation

Gaze labels for turnaround frames generated from single photos, used to train a gaze-conditioned ControlNet that reorients a person's whole body toward a target gaze.

This repository contains labels only — no images. That is deliberate; see Why there are no images here. The frames are regenerated from the pipeline in the code repository, which makes every row here reproducible without redistributing anyone's photographs.

• Code and full reproduction steps: https://github.com/VicsonPeng/GazeCtrl_training-based
• Trained weights: https://huggingface.co/baki0115/gaze-controlnet-qwen-image-edit

What is actually curated here

Automatic gaze estimators can be run by anyone. What cannot be re-derived is the human pass: 888 of the 1864 HITL rows carry a hand-set label, and the SHHQ set went through several rounds of hand culling (1268 kept from a much larger pool, then 121 more removed after a frame-by-frame audit of the head boxes). That curation is the contribution.

Files

file	rows	identities	label sources
labels/hitl_train.csv	1864	183	manual 888 · 3DGazeNet 623 · 6DRepNet 353
labels/hitl_heldout.csv	95	8	6DRepNet 65 · 3DGazeNet 30 — never trained on
labels/shhq_clean.csv	1268	50	6DRepNet 656 · 3DGazeNet 356 · manual 256
labels/shhq_stage1.csv	1147	50	same, minus 121 frames with a bad head box
boxes/head_boxes_hitl.json	1864		head boxes, 512 pad-to-square space
boxes/head_boxes_shhq.json	3433		head boxes for the SHHQ frames
boxes/eye_boxes_hitl.json	1114		eye boxes, for eye-region loss weighting
boxes/headbox_rejects.json	121		frame ids removed in the head-box audit

hitl_* come from COCO and WiderFace source photos (close-up, eyes usually visible). shhq_* come from SHHQ source photos (full body). The two sets share no identities.

Label schema

column	meaning
id	frame id — the key for the box files, and the stem of the regenerated image
sample	identity / source photo it was generated from
mode	turnaround motion: la look-around, tl/tr turn left/right, ud up-down; SHHQ uses back, look, sweep, pitch, pitch_prof
frame	frame index within the generated video
dx, dy, dz	unit gaze vector — dx>0 image-right, dy>0 up, dz>0 toward camera
source	who produced this label: manual, 3DGazeNet, 6DRepNet (also 3dgaze/6drep)
image	path of the regenerated frame, relative to the dataset root

Box files map id → [x1, y1, x2, y2] in the 512×512 pad-to-square space, not in original image coordinates. Crop augmentation must re-map them; the code repository shows how.

Known defects — read before using

Label provenance is confounded with gaze direction. Counting shhq_clean.csv:

source	front	3/4	side	back-3/4	back
3DGazeNet (eye gaze)	174	146	36	0	0
6DRepNet (head pose)	103	132	239	63	119
manual	0	0	242	0	0

3DGazeNet needs visible eyes, so it never labels a profile or a back. 6DRepNet carries those instead, and the human labels sit entirely on profiles. The meaning of a label therefore shifts systematically with the direction it describes — eye gaze at the front, head pose at the sides and back. This is worse than random noise, because it correlates with the variable being modelled. On frames where both estimators fired they disagree by a median 26.7°, which puts a floor under any error metric computed against these labels.

Other things to know:

• hitl_heldout.csv has no human labels at all, so numbers measured on it are pessimistic relative to the training split.
• 6DRepNet's raw dx/dy are sign-flipped against 3DGazeNet, and it flips dx again past profile. The correction is already applied in these files; it is documented in the code repository so you can verify it rather than trust it.
• The frames are generated by Wan2.2 from a single photo. They are plausible, not real captures of that person turning, and generation artefacts survive into the labels.

Why there are no images here

The frames are derivative works of SHHQ, WiderFace and COCO photographs, and they depict real people's faces. SHHQ and WiderFace are distributed for non-commercial research under terms that do not obviously permit redistributing derivatives, so republishing the frames is not mine to do. Regenerate them instead — the pipeline is deterministic given the same source photos, and image/id tell you exactly which frame each row describes.

If you need the frames for a comparison and regeneration is impractical, open an issue on the code repository.

Citation

@misc{gazectrl_labels,
  title  = {GazeCtrl: human-curated gaze labels for body reorientation},
  author = {Peng, Vicson},
  year   = {2026},
  url    = {https://huggingface.co/datasets/baki0115/GazeCtrl_dataset}
}

Licence applies to the labels in this repository. The underlying photographs remain under their original terms.