L2CS-Net (Gaze Estimation)

Model Description

L2CS-Net regresses gaze direction (pitch, yaw) from a face crop. It formulates gaze estimation as a 90-bin classification problem over each axis (4°/bin resolution covering [-180°, +180°]), then computes the expected value across bins for a continuous angle output. ResNet50 backbone, two parallel FC heads.

Reported accuracy (from the original L2CS-Net paper):

• Gaze360 test split: ~3.92° MAE
• MPIIFaceGaze leave-one-out: ~4.16° MAE

These are state-of-the-art numbers for gaze-from-face-crop estimation (2022); for context, geometric iris-eye approaches typically land at 8-15° MAE on the same benchmarks.

Model Details

• Architecture: ResNet50 + dual classification heads (2 × 90-bin)
• Input: 224 × 224 RGB face crop, ImageNet-normalized
• Output: pitch, yaw in radians (head-centric frame)
• Bin resolution: 4° per bin, covering [-180°, +180°]
• Backbones available: ResNet50 (default), ResNet18 (lighter)
• Framework: PyTorch (port of upstream MIT code)

Training data (upstream)

• Gaze360 (Kellnhofer et al., 2019): in-the-wild gaze annotations with 360° head pose coverage. ~127k images.
• MPIIFaceGaze (Zhang et al., 2017): unconstrained office captures with screen-targeted gaze ground truth. ~213k images.

The upstream maintainer trains separate checkpoints for each dataset. Py-Feat exposes the Gaze360 weights as the default since they generalize better to in-the-wild input.

Model Sources

• Original repository (MIT): Ahmednull/L2CS-Net
• Paper: arXiv:2203.03339
• Pretrained weights (upstream): Google Drive folder linked from the upstream README (L2CSNet_gaze360.pkl). Py-Feat hosts a re-packaged .safetensors version on this repo to avoid the pickle (.pkl) deserialization path on user machines. The conversion is documented at scripts/convert_l2cs_pickle_to_safetensors.py in the py-feat repo; no architecture or weight values are modified.

Acknowledgements

This distribution is a re-host of the official L2CS-Net weights trained by Abdelrahman, Hempel, Khalifa, and Al-Hamadi (Otto-von-Guericke University Magdeburg). All training, hyperparameter selection, and benchmark numbers are credited to the original authors. The py-feat project provides only:

• A PyTorch port of the inference code at feat/gaze_detectors/l2cs/l2cs_model.py
• This re-packaged .safetensors artifact for downstream safety
• Integration with feat.Detector and feat.MPDetector's pipelines

Training data are credited to:

• Gaze360 — Kellnhofer, Recasens, Stent, Matusik, Torralba (MIT)
• MPIIFaceGaze — Zhang, Sugano, Fritz, Bulling (MPI Saarbrücken)

Citation

@article{l2csnet2022,
  title={L2CS-Net: Fine-Grained Gaze Estimation in Unconstrained Environments},
  author={Abdelrahman, Ahmed and Hempel, Thorsten and Khalifa, Aly and
          Al-Hamadi, Ayoub},
  journal={arXiv preprint arXiv:2203.03339},
  year={2022}
}

License

MIT — both the original implementation and the converted weights. See upstream LICENSE. Training data licenses (Gaze360, MPIIFaceGaze) are research-use only; commercial deployment may require separate validation.