/laeonetplus

LAEO-Net++

Primary LanguagePython

LAEO-Net++

Try In Colab

Official support code for LAEO-Net++ paper (IEEE TPAMI, 2021).

The LAEO-Net++ architecture
The LAEO-Net++ receives as input two tracks of head crops and a track of maps containing the relative position of the heads, and returns the probability of being LAEO those two heads.

See previous version here

Quick start

The following demo predicts the LAEO label on a pair of heads included in subdirectory data/ava_val_crop. You can choose either to use a model trained on UCO-LAEO or a model trained on AVA-LAEO.

cd laeonetplus
python mains/ln_demo_test.py

Training code will be available soon.

LAEO prediction on a given input video file

Try this Google Colab to run LAEO prediction on a target video file: notebook

Detecting and tracking heads

Try this Google Colab to perform head detection and tracking on a target video file: notebook

Install

Clone this repository:

git clone git@github.com:AVAuco/laeonetplus.git

See library dependencies here. In addition, a requirements.txt file is provided.

Training (in progress)

We provide an example of training on AVA-LAEO data (using preprocessed samples): see the training document.

References

Marin-Jimenez, M. J., Kalogeiton, V., Medina-Suarez, P., Zisserman, A. (2021). LAEO-Net++: revisiting people Looking At Each Other in videos. IEEE transactions on Pattern Analysis and Machine Intelligence, PP, 10.1109/TPAMI.2020.3048482. Advance online publication. https://doi.org/10.1109/TPAMI.2020.3048482

@article{marin21pami,
  author    = {Mar\'in-Jim\'enez, Manuel J. and Kalogeiton, Vicky and Medina-Su\'arez, Pablo and and Zisserman, Andrew},
  title     = {{LAEO-Net++}: revisiting people {Looking At Each Other} in videos},
  journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  doi       = {10.1109/TPAMI.2020.3048482},
  year      = {2021}
}

Conference version:

@inproceedings{marin19cvpr,
  author    = {Mar\'in-Jim\'enez, Manuel J. and Kalogeiton, Vicky and Medina-Su\'arez, Pablo and and Zisserman, Andrew},
  title     = {{LAEO-Net}: revisiting people {Looking At Each Other} in videos},
  booktitle = CVPR,
  year      = {2019}
}