Evaluation of a Monocular Eye Tracking Set-Up

Overview

Evaluation of a Monocular Eye Tracking Set-Up

As part of my master thesis, I implemented a new state-of-the-art model that is based on the work of Chen et al..
For 9 calibration samples, the previous state-of-the-art performance can be improved by up to 5.44% (2.553 degrees compared to 2.7 degrees) and for 128 calibration samples, by 7% (2.418 degrees compared to 2.6 degrees). This is accomplished by (a) improving the extraction of eye features, (b) refining the fusion process of these features, (c) removing erroneous data from the MPIIFaceGaze dataset during training, and (d) optimizing the calibration method.

A software to collect own gaze data and the full gaze tracking pipeline is also available.

Results of the different models.

For the citaitions [1] - [10] please see below. "own model 1" represents the model described in the section below. "own model 2" uses the same model architecture as "own model 1" but is trained without the erroneous data, see MPIIFaceGaze section below. "own model 3" is the same as "own model 2" but with the calibrations points organized in a $\sqrt{k}\times\sqrt{k}$ grid instead of randomly on the screen.

Model

Since the feature extractors share the same weights for both eyes, it has been shown experimentally that the feature extraction process can be improved by flipping one of the eye images so that the noses of all eye images are on the same side. The main reason for this is that the images of the two eyes are more similar this way and the feature extractor can focus more on the relevant features, rather than the unimportant features, of either the left or the right eye.

The architectural improvement that has had the most impact is the improved feature fusion process of left and right eye features. Instead of simply combining the two features, they are combined using Squeeze-and-Excitation (SE) blocks. This introduces a control mechanism for the channel relationships of the extracted feature maps that the model can learn serially.

Start training by running python train.py --path_to_data=./data --validate_on_person=1 --test_on_person=0. For pretrained models, please see evaluation section.

Data

While examining and analyzing the most commonly used gaze prediction dataset, MPIIFaceGaze a subset of MPIIGaze, in detail. It was realized that some recorded data does not match the provided screen sizes. For participant 2, 7, and 10, 0.043%, 8.79%, and 0.39% of the gazes directed at the screen did not match the screen provided, respectively. The left figure below shows recorded points in the datasets that do not match the provided screen size. These false target gaze positions are also visible in the right figure below, where the gaze point that are not on the screen have a different yaw offset to the ground truth.

Results of the MPIIFaceGaze analysis

To the best of our knowledge, we are the first to address this problem of this widespread dataset, and we propose to remove all days with any errors for people 2, 7, and 10, resulting in a new dataset we call MPIIFaceGaze-. This would only reduce the dataset by about 3.2%. As shown in the first figure, see "own model 2", removing these erroneous data improves the model's overall performance.

For preprocessing MPIIFaceGaze, download the original dataset and then run python dataset/mpii_face_gaze_preprocessing.py --input_path=./MPIIFaceGaze --output_path=./data. Or download the preprocessed dataset.

To only generate the CSV files with all filenames which gaze is not on the screen, run python dataset/mpii_face_gaze_errors.py --input_path=./MPIIFaceGaze --output_path=./data. This can be run on MPIIGaze and MPIIFaceGaze, or the CSV files can be directly downloaded for MPIIGaze and MPIIFaceGaze.

Calibration

Nine calibration samples has become the norm for the comparison of different model architectures using MPIIFaceGaze. When the calibration points are organized in a $\sqrt{k}\times\sqrt{k}$ grid instead of randomly on the screen, or all in one position, the resulting person-specific calibration is more accurate. The three different ways to distribute the calibration point are compared in the figure below, also see "own model 3" in the first figure. Nine calibration samples aligned in a grid result in a lower angular error than 9 randomly positioned calibration samples.

To collect your own calibration data or dataset, please refer to gaze data collection.

Comparison of the position of the calibration samples.

Evaluation

For evaluation, the trained models are evaluated on the full MPIIFaceGaze, including the erroneous data, for a fair comparison to other approaches. Download the pretrained "own model 2" models and run python eval.py --path_to_checkpoints=./pretrained_models --path_to_data=./data to reproduce the results shown in the figure above and the table below. --grid_calibration_samples=True takes a long time to evaluate, for the ease of use the number of calibration runs is reduced to 500.

random calibration
k=9
random calibration
k=128
grid calibration
k=9
grid calibration
k=128

k=all
p00 1.780 1.676 1.760 1.674 1.668
p01 1.899 1.777 1.893 1.769 1.767
p02 1.910 1.790 1.875 1.787 1.780
p03 2.924 2.729 2.929 2.712 2.714
p04 2.355 2.239 2.346 2.229 2.229
p05 1.836 1.720 1.826 1.721 1.711
p06 2.569 2.464 2.596 2.460 2.455
p07 3.823 3.599 3.737 3.562 3.582
p08 3.778 3.508 3.637 3.501 3.484
p09 2.695 2.528 2.667 2.526 2.515
p10 3.241 3.126 3.199 3.105 3.118
p11 2.668 2.535 2.667 2.536 2.524
p12 2.204 1.877 2.131 1.882 1.848
p13 2.914 2.753 2.859 2.754 2.741
p14 2.161 2.010 2.172 2.052 1.998
mean 2.584 2.422 2.553 2.418 2.409

Bibliography

[1] Zhaokang Chen and Bertram E. Shi, “Appearance-based gaze estimation using dilated-convolutions”, Lecture Notes in Computer Science, vol. 11366, C. V. Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, Eds., pp. 309–324, 2018. DOI: 10.1007/978-3-030-20876-9_20. [Online]. Available: https://doi.org/10.1007/978-3-030-20876-9_20.
[2] ——, “Offset calibration for appearance-based gaze estimation via gaze decomposition”, in IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, IEEE, 2020, pp. 259–268. DOI: 10.1109/WACV45572.2020.9093419. [Online]. Available: https://doi.org/10.1109/WACV45572.2020.9093419.
[3] Tobias Fischer, Hyung Jin Chang, and Yiannis Demiris, “RT-GENE: real-time eye gaze estimation in natural environments”, in Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part X, Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, Eds., ser. Lecture Notes in Computer Science, vol. 11214, Springer, 2018, pp. 339–357. DOI: 10.1007/978-3-030-01249-6_21. [Online]. Available: https://doi.org/10.1007/978-3-030-01249-6_21.
[4] Erik Lindén, Jonas Sjöstrand, and Alexandre Proutière, “Learning to personalize in appearance-based gaze tracking”, pp. 1140–1148, 2019. DOI: 10.1109/ICCVW.2019.00145. [Online]. Available: https://doi.org/10.1109/ICCVW.2019.00145.
[5] Gang Liu, Yu Yu, Kenneth Alberto Funes Mora, and Jean-Marc Odobez, “A differential approach for gaze estimation with calibration”, in British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018, BMVA Press, 2018, p. 235. [Online]. Available: http://bmvc2018.org/contents/papers/0792.pdf.
[6] Seonwook Park, Shalini De Mello, Pavlo Molchanov, Umar Iqbal, Otmar Hilliges, and Jan Kautz, “Few-shot adaptive gaze estimation”, pp. 9367–9376, 2019. DOI: 10.1109/ICCV.2019.00946. [Online]. Available: https://doi.org/10.1109/ICCV.2019.00946.
[7] Seonwook Park, Xucong Zhang, Andreas Bulling, and Otmar Hilliges, “Learning to find eye region landmarks for remote gaze estimation in unconstrained settings”, Bonita Sharif and Krzysztof Krejtz, Eds., 21:1–21:10, 2018. DOI: 10.1145/3204493.3204545. [Online]. Available: https://doi.org/10.1145/3204493.3204545.
[8] Yu Yu, Gang Liu, and Jean-Marc Odobez, “Improving few-shot user-specific gaze adaptation via gaze redirection synthesis”, pp. 11 937–11 946, 2019. DOI: 10.1109/CVPR.2019.01221. [Online]. Available: http://openaccess.thecvf.com/content_CVPR_2019/html/Yu_Improving_Few-Shot_User-Specific_Gaze_Adaptation_via_Gaze_Redirection_Synthesis_CVPR_2019_paper.html.
[9] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling, “It’s written all over your face: Full-face appearance-based gaze estimation”, pp. 2299–2308, 2017. DOI: 10.1109/CVPRW.2017.284. [Online]. Available: https://doi.org/10.1109/CVPRW.2017.284
[10] ——, “Mpiigaze: Real-world dataset and deep appearance-based gaze estimation”, IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, no. 1, pp. 162–175, 2019. DOI: 10.1109/TPAMI.2017.2778103. [Online]. Available: https://doi.org/10.1109/TPAMI.2017.2778103. \

Owner
Pascal
Pascal
Finding project directories in Python (data science) projects, just like there R rprojroot and here packages

Find relative paths from a project root directory Finding project directories in Python (data science) projects, just like there R here and rprojroot

Daniel Chen 102 Nov 16, 2022
.npy, .npz, .mtx converter.

npy-converter Matrix Data Converter. Expand matrix for multi-thread, multi-process Divid matrix for multi-thread, multi-process Support: .mtx, .npy, .

taka 1 Feb 07, 2022
A collection of learning outcomes data analysis using Python and SQL, from DQLab.

Data Analyst with PYTHON Data Analyst berperan dalam menghasilkan analisa data serta mempresentasikan insight untuk membantu proses pengambilan keputu

6 Oct 11, 2022
Pipeline and Dataset helpers for complex algorithm evaluation.

tpcp - Tiny Pipelines for Complex Problems A generic way to build object-oriented datasets and algorithm pipelines and tools to evaluate them pip inst

Machine Learning and Data Analytics Lab FAU 3 Dec 07, 2022
PySpark Structured Streaming ROS Kafka ApacheSpark Cassandra

PySpark-Structured-Streaming-ROS-Kafka-ApacheSpark-Cassandra The purpose of this project is to demonstrate a structured streaming pipeline with Apache

Zekeriyya Demirci 5 Nov 13, 2022
MotorcycleParts DataAnalysis python

We work with the accounting department of a company that sells motorcycle parts. The company operates three warehouses in a large metropolitan area.

NASEEM A P 1 Jan 12, 2022
Important dataframe statistics with a single command

quick_eda Receiving dataframe statistics with one command Project description A python package for Data Scientists, Students, ML Engineers and anyone

Sven Eschlbeck 2 Dec 19, 2021
ped-crash-techvol: Texas Ped Crash Tech Volume Pack

ped-crash-techvol: Texas Ped Crash Tech Volume Pack In conjunction with the Final Report "Identifying Risk Factors that Lead to Increase in Fatal Pede

Network Modeling Center; Center for Transportation Research; The University of Texas at Austin 2 Sep 28, 2022
MeSH2Matrix - A set of Python codes for the generation of biomedical ontologies from the MeSH keywords of the PubMed scholarly publications

A set of Python codes for the generation of biomedical ontologies from the MeSH keywords of the PubMed scholarly publications

SisonkeBiotik 6 Nov 30, 2022
This repository contains some analysis of possible nerdle answers

Nerdle Analysis https://nerdlegame.com/ This repository contains some analysis of possible nerdle answers. Here's a quick overview: nerdle.py contains

0 Dec 16, 2022
First steps with Python in Life Sciences

First steps with Python in Life Sciences This course material is part of the "First Steps with Python in Life Science" three-day course of SIB-trainin

SIB Swiss Institute of Bioinformatics 22 Jan 08, 2023
Data Competition: automated systems that can detect whether people are not wearing masks or are wearing masks incorrectly

Table of contents Introduction Dataset Model & Metrics How to Run Quickstart Install Training Evaluation Detection DATA COMPETITION The COVID-19 pande

Thanh Dat Vu 1 Feb 27, 2022
Analysis of a dataset of 10000 passwords to find common trends and mistakes people generally make while setting up a password.

Analysis of a dataset of 10000 passwords to find common trends and mistakes people generally make while setting up a password.

Aryan Raj 7 Sep 04, 2022
AptaMat is a simple script which aims to measure differences between DNA or RNA secondary structures.

AptaMAT Purpose AptaMat is a simple script which aims to measure differences between DNA or RNA secondary structures. The method is based on the compa

GEC UTC 3 Nov 03, 2022
Full automated data pipeline using docker images

Create postgres tables from CSV files This first section is only relate to creating tables from CSV files using postgres container alone. Just one of

1 Nov 21, 2021
Elasticsearch tool for easily collecting and batch inserting Python data and pandas DataFrames

ElasticBatch Elasticsearch buffer for collecting and batch inserting Python data and pandas DataFrames Overview ElasticBatch makes it easy to efficien

Dan Kaslovsky 21 Mar 16, 2022
Deep universal probabilistic programming with Python and PyTorch

Getting Started | Documentation | Community | Contributing Pyro is a flexible, scalable deep probabilistic programming library built on PyTorch. Notab

7.7k Dec 30, 2022
Demonstrate a Dataflow pipeline that saves data from an API into BigQuery table

Overview dataflow-mvp provides a basic example pipeline that pulls data from an API and writes it to a BigQuery table using GCP's Dataflow (i.e., Apac

Chris Carbonell 1 Dec 03, 2021
A set of functions and analysis classes for solvation structure analysis

SolvationAnalysis The macroscopic behavior of a liquid is determined by its microscopic structure. For ionic systems, like batteries and many enzymes,

MDAnalysis 19 Nov 24, 2022
Python data processing, analysis, visualization, and data operations

Python This is a Python data processing, analysis, visualization and data operations of the source code warehouse, book ISBN: 9787115527592 Descriptio

FangWei 1 Jan 16, 2022