IrohXu/Gaze-Co-Benchmark

Gaze-Co Benchmark

Evaluation set for Promptable Gaze Target Estimation (PGE), the task introduced in Gaze Target Estimation Anywhere with Concepts (CVPR 2026).

Each instance pairs an image with a natural-language description of one person and the gaze target that person is looking at. A model receives the image and the text prompt — no head bounding box, no detector, no pose — and must localize the subject and predict where they are looking.

The benchmark re-annotates the test splits of three public gaze datasets with concept prompts covering appearance, spatial position, action, and body pose.

Splits

Config	Instances	Images	Gaze points	In-frame	Images included	Source dataset
gazefollow	2,436	2,436	22,895	2,436	✅	GazeFollow
videoattentiontarget	5,301	2,350	5,301	3,417	✅	VideoAttentionTarget
childplay	1,238	536	1,238	1,056	❌ annotations only	ChildPlay

GazeFollow ships 6–19 independent annotator points per instance; all of them are kept in gaze_all so the standard GazeFollow protocol can be reproduced. The other two splits have a single annotator per instance and include out-of-frame targets.

ChildPlay images are not redistributed here. Its parquet shards carry every annotation column but no image. Obtain the frames from the ChildPlay authors under their data agreement, then place them under images/<image_path> (see below).

Seven GazeFollow instances whose four concept prompts were all empty in the source annotations are excluded: with no prompt there is nothing to condition on, so they are not scorable under PGE. The published split is therefore 2,436, not 2,443.

Loading

from datasets import load_dataset

ds = load_dataset("IrohXu/Gaze-Co-Benchmark", "gazefollow", split="test")
row = ds[0]
row["image"]        # PIL.Image  (absent in the childplay config)
row["appearance"]   # "short brown hair, wearing a plaid shirt and red pants man"
row["gaze_all"]     # [[0.284, 0.86], [0.294, 0.793], ...] normalized xy, one per annotator
row["head_bbox"]    # [x1, y1, x2, y2] normalized — reference only, never a model input
row["inout"]        # 1 if the gaze target is inside the frame

Fields

Field	Type	Description
image	Image	The RGB frame. Present only for gazefollow and videoattentiontarget.
image_path	string	Path relative to the source dataset root, e.g. gazefollow/test2/00000000/00000001.jpg.
source	string	gazefollow, videoattentiontarget or childplay.
instance_id	int32	1-based index within the split.
width, height	int32	Image size in pixels.
head_bbox	list[float] (4)	Head box of the prompted person, normalized xyxy. Ground truth for the localization head; not a model input under PGE.
gaze	list[float] (2)	Mean of all annotator points, normalized xy.
gaze_all	list[list[float]]	Every annotator point, normalized xy.
num_annotators	int32	Length of gaze_all.
inout	int32	1 = target inside the frame, 0 = outside. When 0, GazeFollow/VideoAttentionTarget store (-1, -1) and ChildPlay stores a direction point; gate on this field.
appearance	string	Appearance prompt — the default text input for GazeAnywhere.
position	string	Spatial-position prompt, e.g. "left side of the image".
action	string	Action prompt, e.g. "using a white laptop".
pose	string	Body-pose prompt, e.g. "sitting on a couch and looking down".
meta	string	Source-specific provenance (sub-collection, clip id, original resolution).

Evaluation

Metrics follow the protocol of each source dataset:

Split	Metrics
gazefollow	AUC, minimum L2 distance over annotators, L2 distance to the annotator mean
videoattentiontarget	AUC, L2 distance, in/out average precision
childplay	AUC, L2 distance, in/out average precision

Distances are in normalized image coordinates. AUC for GazeFollow is computed at the original image resolution against all annotator points; for the other two splits it is computed at 64×64 against the Gaussian support of the single ground-truth point. AUC and distance cover in-frame targets only; AP covers every instance.

Reproducing

git clone https://github.com/IrohXu/GazeAnywhere && cd GazeAnywhere

# unpack the bundled images once (only pyarrow required)
python tools/prepare_gaze_co.py --benchmark_root /path/to/Gaze-Co-Benchmark

python tools/eval_on_gaze_co.py \
  --config_file    configs/gazeanywhere_config.py \
  --model_weights  /path/to/GazeAnywhere/model.safetensors \
  --benchmark_root /path/to/Gaze-Co-Benchmark \
  --use_dark_inference

prepare_gaze_co.py writes <benchmark_root>/images/ for the two splits that ship pixels and reports ChildPlay as skipped. To score ChildPlay, copy its frames into that same tree so their relative paths match image_path. If you already hold the source datasets, skip the unpacking step entirely and pass --image_root /your/gaze_datasets instead.

--splits restricts which splits run, and --prompt_field {appearance,position,action,pose,none} selects which concept conditions the model (none scores with an empty prompt, which measures how much the text conditioning contributes).

Repository layout

annotations/                     # native GazeAnywhere schema, consumed by GazeAnywhereDataset
  gazefollow_test.json
  videoattentiontarget_test.json
  childplay_test.json
  gaze_co_test.json              # all three splits merged, idx renumbered
  stats.json
data/<source>/test-*.parquet     # Hugging Face shards; image bytes embedded except childplay

The parquet shards are the single copy of the pixel data — there is no duplicate images/ tree in this repository. tools/prepare_gaze_co.py materializes one locally.

Licensing and intended use

The concept prompts, the repackaging, and the evaluation protocol are released under the GazeAnywhere license (see LICENSE) for non-commercial research use.

The underlying images and gaze annotations remain governed by their original licenses and terms of use. By using this benchmark you agree to comply with the terms of every source dataset you touch:

• GazeFollow — images redistributed here; see the original release for its terms.
• VideoAttentionTarget — images redistributed here; see the original release for its terms.
• ChildPlay — images are not redistributed. Request them from the ChildPlay authors and accept their data agreement.

If you represent a source dataset and object to this redistribution, please open a discussion on this repository and we will remove the affected split.

This benchmark is for evaluating gaze target estimation research. It is not validated for clinical, surveillance, biometric identification, or any decision-making use involving individuals.

Citation

@inproceedings{cao2026gaze,
  title={Gaze Target Estimation Anywhere with Concepts},
  author={Cao, Xu and Yang, Houze and Gunda, Vipin and Zhou, Zhongyi and Xu, Tianyu and Kowdle, Adarsh and Kim, Inki and Rehg, James M},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={31304--31315},
  year={2026}
}

Please also cite GazeFollow, VideoAttentionTarget, and ChildPlay when reporting results on the corresponding splits.