Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Improving Robustness of Object Detectors Against Domain Shift

Le-Anh Tran, Chung Nguyen Tran, Dong-Chul Park, Jordi Carrabina, David Castells-Rufas

TL;DR

The paper tackles domain shift in driving-object detection by introducing a data-synthesis module that creates realistic hazy images. It leverages Monodepth2 for monocular depth estimation and Koschmieder’s law to render hazy scenes with random haze density and airlight, enabling online or offline data augmentation for YOLOv4-based detectors. The approach yields substantial improvements on hazy and foggy data (e.g., WAYMO-Haze, Foggy Driving) and provides modest gains on clean data, highlighting enhanced robustness to adverse weather. This method offers a practical, label-efficient route to strengthen autonomous-driving perception in varying visibility conditions, with publicly available code for replication.

Abstract

This paper proposes a data augmentation method for improving the robustness of driving object detectors against domain shift. Domain shift problem arises when there is a significant change between the distribution of the source data domain used in the training phase and that of the target data domain in the deployment phase. Domain shift is known as one of the most popular reasons resulting in the considerable drop in the performance of deep neural network models. In order to address this problem, one effective approach is to increase the diversity of training data. To this end, we propose a data synthesis module that can be utilized to train more robust and effective object detectors. By adopting YOLOv4 as a base object detector, we have witnessed a remarkable improvement in performance on both the source and target domain data. The code of this work is publicly available at https://github.com/tranleanh/haze-synthesis.

Toward Improving Robustness of Object Detectors Against Domain Shift

TL;DR

The paper tackles domain shift in driving-object detection by introducing a data-synthesis module that creates realistic hazy images. It leverages Monodepth2 for monocular depth estimation and Koschmieder’s law to render hazy scenes with random haze density and airlight, enabling online or offline data augmentation for YOLOv4-based detectors. The approach yields substantial improvements on hazy and foggy data (e.g., WAYMO-Haze, Foggy Driving) and provides modest gains on clean data, highlighting enhanced robustness to adverse weather. This method offers a practical, label-efficient route to strengthen autonomous-driving perception in varying visibility conditions, with publicly available code for replication.

Abstract

This paper proposes a data augmentation method for improving the robustness of driving object detectors against domain shift. Domain shift problem arises when there is a significant change between the distribution of the source data domain used in the training phase and that of the target data domain in the deployment phase. Domain shift is known as one of the most popular reasons resulting in the considerable drop in the performance of deep neural network models. In order to address this problem, one effective approach is to increase the diversity of training data. To this end, we propose a data synthesis module that can be utilized to train more robust and effective object detectors. By adopting YOLOv4 as a base object detector, we have witnessed a remarkable improvement in performance on both the source and target domain data. The code of this work is publicly available at https://github.com/tranleanh/haze-synthesis.
Paper Structure (10 sections, 2 equations, 6 figures, 1 table)

This paper contains 10 sections, 2 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Haze Imaging Model
  4. Depth Estimation Model
  5. Object Detection Model
  6. Methodology
  7. Experiments
  8. Synthetic Image Data
  9. Object Detection Performance
  10. Conclusions

Figures (6)

  • Figure 1: Data synthesis: (a) Original clean image, (b) Synthetic image by the proposed method, (c) Natural hazy image sakaridis2018semantic.
  • Figure 2: Pipeline of the proposed method (the illustration of YOLOv4 is adapted from cai2021yolov4).
  • Figure 3: Synthetic image data: (a) Original clean image, (b) Estimated depth, (c,d,e) Synthetic hazy images by the proposed method with different $\beta$ values: 1.0, 1.5, 2.0, respectively, note that larger value of $\beta$ results in denser haze, and (f) Synthetic hazy image based on transmission randomization.
  • Figure 4: Data synthesis results for KITTI (three top pairs) and MSCOCO image data (left: original, right: synthetic).
  • Figure 5: Object detection performance comparison: (a) YOLOv4 on clean scene, (b) YOLOv4 on hazy scene, (c) YOLOv4-Synt on clean scene, and (d) YOLOv4-Synt on hazy scene.
  • ...and 1 more figures