ChildPlay: A New Benchmark for Understanding Children's Gaze Behaviour
Samy Tafasca *, Anshul Gupta *, Jean-Marc Odobez (* equal contribution)
ICCV 2023
[Paper] [Video] [Dataset]

This repository provides the official code and checkpoints for the GeomGaze model, as introduced in our ICCV paper, ChildPlay: A New Benchmark for Understanding Children's Gaze Behaviour. It also includes annotations and scripts for our novel semantic metric that evaluates gaze performance when looking at heads.

The GeomGaze model constructs a geometrically consistent point cloud of the scene. This point cloud is matched with a predicted 3D gaze vector to compute the 3D Field-of-View (3DFoV), highlighting visible regions in 3D. The 3DFoV is then combined with the scene image to predict the final gaze target.

Download the required datasets:

• GazeFollow extended: [Download]
• VideoAttentionTarget: [Download]
• ChildPlay: [Download]

Update the dataset paths (*_data) in config.py accordingly.

Additionally, download our processed data: [Download]

• Validation labels: found in labels/ after extraction from the download.
  • Update *_train_label, *_val_label, and *_test_label in the config.
• Fixed image cropping parameters: found in val_crop_params/ after extraction from the download.
  • Update *_val_crop_params in the config.

1. Extract Depth Maps:

   • Use the SamsungLabs depth estimation model with domain=depth.
   • We use the b5_lrn4 model.

2. Extract Focal Length:

   • Use the AdelaiDepth model with ResNeXt101 backbone.
   • Save focal lengths as separate .txt files per image.
   • We provide a modified inference script at utils/test_shape.py.
   • Optionally approximate focal length with the longest side of the image in pixels (there may be a loss in performance).

Ensure the extracted outputs follow the dataset directory structure and update *_depth and *_focal_length in the config.

We use PyTorch for our experiments. Install dependencies using:

conda env create -f environment.yml

python train.py --dataset GazeFollow

python train.py --dataset VideoAtt --init_weights <path>

Provide initial weights from training on GazeFollow using --init_weights.

python train.py --dataset ChildPlay --init_weights <path>

Provide initial weights from training on GazeFollow using --init_weights.

python test_on_gazefollow.py --orig_ar --model_weights <path> --csv_path <csv_path>

Provide the model weights using --model_weights and the output path for predictions using --csv_path.

python eval_on_vat_childplay.py --orig_ar --model_weights <path> --dataset <dataset> --csv_path <csv_path>

Specify the dataset (ChildPlay or VideoAtt) using --dataset.

1. Download our annotations: found in LAH_annotations/ after extraction from the download.
2. Update bbox_path and gt_path in compute_lah.py. The data_path remains as per config.py.
3. Also update the dataset, subset (only for ChildPlay) and pred_path to the predictions csv.
4. Compute the LAH scores:

python compute_lah.py

Our checkpoints are available under the same download link as our processed data: [Download]

If you use our code, please cite:

@InProceedings{Tafasca_2023_ICCV,
    author    = {Tafasca*, Samy and Gupta*, Anshul and Odobez, Jean-Marc},
    title     = {ChildPlay: A New Benchmark for Understanding Children's Gaze Behaviour},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {20935-20946},
    note      = {* Equal contribution}
}

This code is adapted from our previous work:

• idiap/multimodal_gaze_target_prediction
  • This work, in turn, leverages code from ejcgt/attention-target-detection.

We thank the authors for their contributions.