Toward Robust Pedestrian Detection With Data Augmentation

Sebastian Cygert,Andrzej Czyzewski

doi:10.1109/access.2020.3011356

Sebastian Cygert, Andrzej Czyzewski

Open Access

https://doi.org/10.1109/access.2020.3011356

Copy DOI

Abstract

In this article, the problem of creating a safe pedestrian detection model that can operate in the real world is tackled. While recent advances have led to significantly improved detection accuracy on various benchmarks, existing deep learning models are vulnerable to invisible to the human eye changes in the input image which raises concerns about its safety. A popular and simple technique for improving robustness is using data augmentation. In this work, the robustness of existing data augmentation techniques is evaluated to propose a new simple augmentation scheme where during training, an image is combined with a patch of a stylized version of that image. Evaluation of pedestrian detection models robustness and uncertainty calibration under naturally occurring corruption and in realistic cross-dataset evaluation setting is conducted to show that our proposed solution improves upon previous work. In this paper, the importance of testing the robustness of recognition models is emphasized and it shows a simple way to improve it, which is a step towards creating robust pedestrian and object detection models.

Highlights

In recent years visual recognition has witnessed a significant progress, mainly due to the introduction of Convolutional Neural Networks (CNN) and the availability of large scale datasets
In this work, the examination was performed of pedestrian detection models in the real-world setting when test-time data come from a different distribution than in training: using cross-dataset evaluation, testing the model by switching illumination conditions and through testing it on synthetic distortions
We show that data augmentations in the form of stylized and Gaussian augmentations significantly improve the robustness of the model

Summary

INTRODUCTION

In recent years visual recognition has witnessed a significant progress, mainly due to the introduction of Convolutional Neural Networks (CNN) and the availability of large scale datasets. Even though CNN based models surpassed human performance on some of the benchmarks [1], the application of deep learning methods in safety-critical applications like medicine or autonomous-vehicles has been limited [2] This is due to the fact that CNNs often fail to generalize outside of the training data distribution. Test-time distribution of data differs from the training distribution, when deep learning models predict wrong output with high confidence [11] It is of particular importance for models operating in the real-world to be robust to such distributional changes. We analyze the impact of the distributional shift on the accuracy of pedestrian detection models, i.e. detection models are evaluated in cross-dataset setting, by adding different types of image distortions and when testing on night-time images,.

RELATED WORK

EXPERIMENTAL SETUP

EXPERIMENTS

Findings

CONCLUSION