1 of 6

DarkLighter: Light Up the Darkness for UAV Tracking

* Corresponding author

Presenter

Junjie Ye, Changhong Fu*, Guangze Zheng, Ziang Cao, and Bowen Li

{ye.jun.jie, changhongfu}@tongji.edu.cn

Tongji University

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

1

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online

2 of 6

Methodology: DarkLighter

Iterative Retinex model

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

2

…

Map estimation network (ME-Net)

Estimated illumination maps

Estimated noise maps

Original image

Enhanced image

…

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online

Inspired by this, we propose an iterative Retinex model in this work. As the figure shows, DarkLighter dedicates to peel off illumination and noise maps iteratively. Given a low-light image S_zero, DarkLighter erases illumination and noise progressively. After a total of I iterations, the illumination and noise maps are relatively cleaned up, and the final output S_I is regarded as the reflectance R of the image.

Therefore, we need to estimate the noise map and the illumination map in every iteration. We train a map estimation network, that is, the ME-Net, to estimate image-specific illumination and noise maps. Specifically, ME-Net consists of seven convolutional layers as shown in the Figure. Taking the initial low-light image as input, after passing through several convolutional layers, a set of illumination maps and noise maps are obtained for iterative decomposition.

3 of 6

Methodology: DarkLighter

Center-focused lightness loss

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

3

Illumination regularization loss

Noise estimation loss

Color intensity loss

Semantic fidelity loss

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online

To enable ME-Net with robust maps estimation ability, several consumption functions are designed to guide the training process. [loss]

DarkLighter towards lighting up the template patch and the search patch. Since the tracked object is centered at the template patch and generally appears in the central area of the search patch, it is worth paying more attention to enhance the central area. In this loss, the enhanced image is divided into **P** nonoverlapping patches. The average intensity value y of each patch is regularized to approximate a well-illumination value *l*. And W is a weight matrix to emphasize the central area.

The illumination regularization loss is added to constrain the smoothness of each illumination component.

The noise estimation loss is tend to restrain the overall intensity of noise maps.

As for the color intensity loss, it bridges three channels and constrains their intensity to a relatively same level.

Moreover, since the semantic information of images plays a pivotal role in object tracking, the degradation of semantic information in the enhanced image can easily lead to tracking failure. To this end, a semantic fidelity loss is introduced, which constrains the difference between enhanced and original images.

4 of 6

Experiments: Efficacy of DarkLighter for UAV Dark Tracking�

DarkLighter on Different Trackers

Implemented on 4 trackers, including DiMP50, PrDiMP50, SiamAPN, and SiamRPN++
Evaluated on the UAVDark135 benchmark

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

4

Precision and success plots of SOTA trackers with DarkLighter (denoted as DL in the legend) enabled or not.

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online

5 of 6

Experiments: Efficacy of DarkLighter for UAV Dark Tracking�

Comparison of DarkLighter and SOTA Enhancers

Compare the performance with other SOTA enhancers in boosting nighttime tracking, including EnlightenGAN, Zero-DCE, and LIME.

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

5

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online

6 of 6

DarkLighter: Light Up the Darkness for UAV Tracking

Junjie Ye, Changhong Fu*, Guangze Zheng, Ziang Cao, and Bowen Li

Thank you for your attention!

J. Ye, C. Fu, G. Zheng, Z. Cao, and B. Li, "DarkLighter: Light Up the Darkness for UAV Tracking," in IROS 2021

6

Lab web

Personal web

Vision4Robotics

2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

September 27-October 1, 2021 Prague, Czech Republic, Online