gaze-estimation

MobileGaze: Pre-trained mobile nets for Gaze-Estimation

Video by Yan Krukau: https://www.pexels.com/video/male-teacher-with-his-students-8617126/

This project aims to perform gaze estimation using several deep learning models like ResNet, MobileNet v2, and MobileOne. It supports both classification and regression for predicting gaze direction. Built on top of L2CS-Net, the project includes additional pre-trained models and refined code for better performance and flexibility.

Features

• ONNX Inference: Export pytorch weights to ONNX and ONNX runtime inference.
• ResNet: Deep Residual Networks - Enables deeper networks with better accuracy through residual learning.
• MobileNet v2: Inverted Residuals and Linear Bottlenecks - Efficient model for mobile applications, balancing performance and computational cost.
• MobileOne (s0-s4): An Improved One millisecond Mobile Backbone - Achieves near-instant inference times, ideal for real-time mobile applications.
• Face Detection: SCFRD - Sample and Computation Redistribution for Efficient Face Detection (SCRFD) model for efficient face detection.

Installation

1. Clone the repository:

git clone https://github.com/yakyo/gaze-estimation.git
cd gaze-estimation

1. Install the required dependencies:

pip install -r requirements.txt

1. Download weight files:

   a) Download weights from the following links:

   Model	Weights	Size	Epochs	MAE
   ResNet-18	resnet18.pt	43 MB	200	12.84
   ResNet-34	resnet34.pt	81.6 MB	200	11.33
   ResNet-50	resnet50.pt	91.3 MB	200	11.34
   MobileNet V2	mobilenetv2.pt	9.59 MB	200	13.07
   MobileOne S0	mobileone_s0_fused.pt	4.8 MB	200	12.58
   MobileOne S1	mobileone_s1_fused.pt	xx MB	200	*
   MobileOne S2	mobileone_s2_fused.pt	xx MB	200	*
   MobileOne S3	mobileone_s3_fused.pt	xx MB	200	*
   MobileOne S4	mobileone_s4_fused.pt	xx MB	200	*

   ’*’ - soon will be uploaded (due to limited computing resources I cannot publish rest of the weights, but you still can train them with given code).

   b) Run the command below to download weights to the weights directory (Linux):

   sh download.sh [model_name]
                  resnet18
                  resnet34
                  resnet50
                  mobilenetv2
                  mobileone_s0
                  mobileone_s1
                  mobileone_s2
                  mobileone_s3
                  mobileone_s4
   

Usage

Datasets

Dataset folder structure:

data/
├── Gaze360/
│   ├── Image/
│   └── Label/
└── MPIIFaceGaze/
    ├── Image/
    └── Label/

Gaze360

• Link to download dataset: https://gaze360.csail.mit.edu/download.php
• Data pre-processing code: https://phi-ai.buaa.edu.cn/Gazehub/3D-dataset/#gaze360

MPIIGaze

• Link to download dataset: download page
• Data pre-processing code: https://phi-ai.buaa.edu.cn/Gazehub/3D-dataset/#mpiifacegaze

Training

python main.py --data [dataset_path] --dataset [dataset_name] --arch [architecture_name]

main.py arguments:

usage: main.py [-h] [--data DATA] [--dataset DATASET] [--output OUTPUT] [--checkpoint CHECKPOINT] [--num-epochs NUM_EPOCHS] [--batch-size BATCH_SIZE] [--arch ARCH] [--alpha ALPHA] [--lr LR] [--num-workers NUM_WORKERS]

Gaze estimation training.

options:
  -h, --help            show this help message and exit
  --data DATA           Directory path for gaze images.
  --dataset DATASET     Dataset name, available `gaze360`, `mpiigaze`.
  --output OUTPUT       Path of output models.
  --checkpoint CHECKPOINT
                        Path to checkpoint for resuming training.
  --num-epochs NUM_EPOCHS
                        Maximum number of training epochs.
  --batch-size BATCH_SIZE
                        Batch size.
  --arch ARCH           Network architecture, currently available: resnet18/34/50, mobilenetv2, mobileone_s0-s4.
  --alpha ALPHA         Regression loss coefficient.
  --lr LR               Base learning rate.
  --num-workers NUM_WORKERS
                        Number of workers for data loading.

Evaluation

python evaluate.py --data [dataset_path] --dataset [dataset_name] --weights [weights_path] --arch [architecture_name]

evaluate.py arguments:

usage: evaluate.py [-h] [--data DATA] [--dataset DATASET] [--weights WEIGHTS] [--batch-size BATCH_SIZE] [--arch ARCH] [--num-workers NUM_WORKERS]

Gaze estimation evaluation.

options:
  -h, --help            show this help message and exit
  --data DATA           Directory path for gaze images.
  --dataset DATASET     Dataset name, available `gaze360`, `mpiigaze`
  --weights WEIGHTS     Path to model weight for evaluation.
  --batch-size BATCH_SIZE
                        Batch size.
  --arch ARCH           Network architecture, currently available: resnet18/34/50, mobilenetv2, mobileone_s0-s4.
  --num-workers NUM_WORKERS
                        Number of workers for data loading.

Inference

detect.py --arch [arch_name] --gaze-weights [path_gaze_estimation_weights] --face-weights [face_det_weights] --view --input [input_file] --output [output_file] --dataset [dataset_name]

detect.py arguments:

usage: detect.py [-h] [--arch ARCH] [--gaze-weights GAZE_WEIGHTS] [--face-weights FACE_WEIGHTS] [--view] [--input INPUT] [--output OUTPUT] [--dataset DATASET]

Gaze Estimation Inference Arguments

options:
  -h, --help            show this help message and exit
  --arch ARCH           Model name, default `resnet18`
  --gaze-weights GAZE_WEIGHTS
                        Path to gaze esimation model weights
  --face-weights FACE_WEIGHTS
                        Path to face detection model weights
  --view                Display the inference results
  --input INPUT         Path to input video file
  --output OUTPUT       Path to save output file
  --dataset DATASET     Dataset name to get dataset related configs

Reference

1. This project is built on top of L2CS-Net. Most of the code parts have been re-written for reproducibility and adaptability. Several additional backbones are provided with pre-trained weights.
2. https://github.com/apple/ml-mobileone
3. https://github.com/yakhyo/face-reidentification (used for inference, modified from insightface)

This site is open source. Improve this page.