FLAME

Original Pytorch Implementation of FLAME: Facial Landmark Heatmap Activated Multimodal Gaze Estimation, accepted at the 17th IEEE Internation Conference on Advanced Video and Signal-based Surveillance, AVSS 2021, Virtual, November 16-19, 2021.

arXiv Preprint: 2110.04828

HAL: hal-03386581v1

Dependencies used

• python v3.6.13
• cuda v10.2
• numpy v1.19.2
• pandas v1.1.3
• opencv-python-headless v4.5.1.48 (or use equivalent opencv)
• torch v1.8.0
• torchvision v0.9.0
• scipy v1.5.2
• matplotlib v3.3.4

FLAME - Model and Results

Model Architecture for FLAME -

Results -

Project Installation

This code requires Python 3, Pytorch, Cuda and other dependencies as listed above. Open Command Prompt/Terminal and use the following commands to clone the project and install dependencies (python, pip are prerequisites) -

$ git clone https://github.com/neelabhsinha/flame.git
$ cd flame
$ pip3 install -r requirements.txt
$ python3 main.py -h

Alternatively you can install the above dependencies in your environment as you prefer. The exact versions used in this project are listed above. To be able to do necessary data preprocessing, you would also optionally require Openface 2.0 (or any other 2D landmark extractor) and RetinaFace (or any other way to extract face crops). These are optional requirements and feel free to use any other toolkit to achieve the required objective. For more details read through the section 'Preparing the dataset'.

Project Description

Files in the project are described in the following table -

Each model as described in the paper is identified with a unique key in this code which we shall address by model_id in this readme. The keys to those models along with the pretrained checkpoints (containing model weights, and other variables related to the final state) are given in the table below -

The format of these checkpoints is in the form of dictionary with following schema -

{
  'epoch': epoch number, 
  'model_state_dict': model_weights, 
  'optimizer_state_dict': optimizer state,
  'scheduler_state_dict': scheduler_state,
  'loss_train': mean_training_loss, 
  'loss_cv': mean_cv_loss
}

Prepraring the dataset

Prepare a directory structure as the following and add the root in the dataset_paths dictionary in config.py.

dataset_root
│
└───images
│   │───subject01
│   │   │   1.npy
│   |   │   2.npy
│   |   │   ...
│   │
│   └───subject02
│   |   │   1.npy
│   |   │   2.npy
│   |   │   ...
│   |....
│   
└───heatmaps
│   │───subject01
│   │   │   1.npy
│   |   │   2.npy
│   |   │   ...
│   │
│   └───subject02
│   |   │   1.npy
│   |   │   2.npy
│   |   │   ...
│   |....
|
└───facial_landmarks_2d
│   │───subject01.pkl
│   |───subject02.pkl
|   |....
│   
|
└───head_pose
│   │───subject01.pkl
│   |───subject02.pkl
|   |....
│   
└───gaze_target
│   │───subject01.pkl
│   |───subject02.pkl
|   |....
│   

• images - RGB images as numpy arrays with dimension (height, width, channels) of the complete face crop. Stored as one folder for each subject, and images of the particular subject contained inside as numpy arrays.
• heatmaps - 2D landmark heatmaps as numpy arrays with dimension (height, width, channels) of the complete face crop. Stored as one folder for each subject, and images of the particular subject contained inside as numpy arrays.
• facial_landmarks_2d - 2D facial landmarks in the form of numpy arrays saved as pickle (index i of a file contains data corresponding to (i+1)th image and heatmap file inside the given subject folder with the same name as that of the pickle file).
• head_pose - Head pose angles in the form of numpy arrays of the type [pitch, yaw, roll] angle numpy arrays (index i of a file contains data corresponding to (i+1)th image and heatmap file inside the given subject folder with the same name as that of the pickle file). If not available, headpose can be used from the output of Openface 2.0.
• gaze_angles - Gaze angles in the form of numpy arrays of the type [yaw, pitch] angle numpy arrays (index i of a file contains data corresponding to (i+1)th image and heatmap file inside the given subject folder with the same name as that of the pickle file). This is the ground truth.

Steps -

1. Extract Face crops from RetinaFace and zero-pad them to nearest 4:5 ratio (or any other way to extract face crops).
2. Crop them to 384 * 450 pixels. The dimension of RGB images, hence, will be 384*450*3.
3. Run Openface 2.0 on these images (or any other way to extract landmarks in desired order as given by Openface)
4. Collect the 2D facial landmarks from it in the above directory structure as per the given instruction above. Each element will be a 112-dimensional numpy array (for 28 2D landmarks for both eyes in the format (fl_x_0, fl_x_1,...,fl_x_55,fl_y_0,fl_y_1,...,fl_y_55), with 0-27 index for the right eye and 28-55 for the left eye. There will be same number of such elements as the number of images in corresponding images folder. So, for a subject, this will be a numpy array of dimensions (num_images, 112).
5. Collect the images, head pose, and gaze targets in the above directory structure as per given instructions. To generate head_pose angles from rotation matrix, use get_head_pose_angles from utils/preprocess.py. For a subject, dimension of head pose will be (number of images, 3) and gaze targets will be (number of images, 2), both as numpy arrays.
6. Add the root dataset directory to dataset_paths by dataset_name:dataset_path in config.py (Use this dataset_name everywhere in the code for all dataset name related parameters in the code).
7. Generate heatmaps from the 2D landmarks after completing step 1-6. You can use the function get_and_save_heatmap given in utils/preprocess.py with dataset_name as parameter. Use the following command -

$ python3 main.py --get_heatmap --dataset <dataset_name>

9. It should create heatmaps directory and save the heatmaps there, with same format as that of images, and each heatmap to be 384*480*28. Corresponding landmarks for left and the right eye will be in the same channel, and one will be cropped out during loading the data when eye crops are generated. This is done to make the heatmap less sparse and save memory during storage, while also make data loading efficient.

The dataset directory is now ready.

Notes

• Maintain 1...n continuous numbers for images, heatmaps and save other data in pickle at corresponding 0-(n-1) indices
• Take care of the file formats

Other Configurations required

Please do the following before running the code

1. Please add all the dependencies in your environment which support the given version.
2. In config.py file, change/add all the dataset paths, and other parameters as defined.

Optional Configurations

Few other metadata that are required but is already given along with this repository for our experiments are described below. You may run it on your own but it's not compulsory.

1. Generating Split - Decide which folders will be in train, test, and val splits. Can be done using the following script (our split is available in metadata/splits directory) -

$ python3 main.py --split_data --split_nature cross-person --data <dataset_name>

Function is available at utils/data_split.py for viewing the schema of the file

2. Getting maximum and minimum values of each input and output - Used for normalization purposes and is extracted only for the training data. For our split and data, the parameters are given in metadata/data_statistics in the form of a dictionary stored as pickle. Use the following command to extract these parameters -

$ python3 main.py --get_data_stats --dataset <dataset_name>

Function is available at utils/preprocess.py by the name get_mean_and_std

Evaluations/Testing

1. Set up the project using above steps
2. Download the weights from the specified locations
3. Execute the following command -

$ python3 main.py --test <model_id> --dataset <dataset_name of training dataset> --test_data <dataset_name of testing dataset> --load_checkpoint <complete path of the checkpoint on which model is to be tested>

A csv file will be stored at the test_path location as specified in config.py by the name '<train_dataset_name>_<test_dataset_name>_<model_id>.csv' having the following schema for all provided images in order -

index, yaw_p, pitch_p, yaw_t, pitch_t, loss_3d, error_y, error_p

Apart from this, a log file is maintained by the name testing_<dataset_name>_<model_name>_cross-person.log in the loggers_loc directory.

Note -

• To generate predictions on the customized pipeline, you can create an input pipeline on your own and use the function forward_propagation inside utils/train.py and provide the inputs to the same. It will return you the values in order of a tuple ((predicted_yaw, predicted_pitch),(true_yaw, true_pitch), error) of type (tensor, tensor, float).

Training

1. Set up the project using above steps
2. Execute the following command -

$ python3 main.py --train <model_id> --dataset <dataset_name>

To change training hyperparameters, change variables in config.py file

Training from a checkpoint -

$ python3 main.py --train <model_id> --dataset <dataset_name> --load_checkpoint <complete path of checkpoint file>

A log file is maintained by the name training_<dataset_name>_<model_name>_cross-person.log in the loggers_loc directory logging details of losses after each epoch is completed.

Citation

If you found our work helpful in your use case, please cite the following paper -

@INPROCEEDINGS{9663816,
  author={Sinha, Neelabh and Balazia, Michal and Bremond, François},  
  booktitle={2021 17th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS)},
  title={FLAME: Facial Landmark Heatmap Activated Multimodal Gaze Estimation},
  year={2021},
  volume={},
  number={},
  pages={1-8},
  doi={10.1109/AVSS52988.2021.9663816}
}