UG2+ Challenge

Bridging the Gap Between Computational Photography and Visual Recognition:

5^th UG²+ Prize Challenge

CVPR 2022

$10K prizes

Location

The rapid development of computer vision algorithms increasingly allows automatic visual recognition to be incorporated into a suite of emerging applications. Some of these applications have less-than-ideal circumstances such as low-visibility environments, causing image captures to have degradations. In other more extreme applications, such as imagers for flexible wearables, smart clothing sensors, ultra-thin headset cameras, implantable in vivo imaging, and others, standard camera systems cannot even be deployed, requiring new types of imaging devices. Computational photography addresses the concerns above by designing new computational techniques and incorporating them into the image capture and formation pipeline. This raises a set of new questions. For example, what is the current state-of-the-art for image restoration for images captured in non-ideal circumstances? How can inference be performed on novel kinds of computational photography devices?

Continuing the success of the 1^st (CVPR'18), 2^nd (CVPR'19), 3^rd (CVPR'20), and 4^th (CVPR'21) UG² Prize Challenge workshops, we provide its 5^th version for CVPR 2021. It will inherit the successful benchmark dataset, platform and evaluation tools used by the first three UG2 workshops, but will also look at brand new aspects of the overall problem, significantly augmenting its existing scope.

Original high-quality contributions are solicited on the following topics:

Novel algorithms for robust object detection, segmentation or recognition on outdoor mobility platforms (UAVs, gliders, autonomous cars, outdoor robots etc.), under real-world adverse conditions and image degradations (haze, rain, snow, hail, dust, underwater, low-illumination, low resolution, etc.)

Novel models and theories for explaining, quantifying, and optimizing the mutual influence between the low-level computational photography tasks and various high-level computer vision tasks, and for the underlying degradation and recovery processes, of real-world images going through complicated adverse visual conditions.

Novel evaluation methods and metrics for image restoration and enhancement algorithms, with a particular emphasis on no-reference metrics, since for most real outdoor images with adverse visual conditions it is hard to obtain any clean “ground truth” to compare with.

Available Challenges

What is the current state-of-the-art for image restoration for images captured in non-ideal circumstances? How can inference be performed on novel kinds of computational photography devices?

The UG²+ Challenge seeks to advance the analysis of "difficult" imagery by applying image restoration and enhancement algorithms to improve analysis performance. Participants are tasked with developing novel algorithms to improve the analysis of imagery captured under problematic conditions.

Object Detection in Haze

While most current vision systems are designed to perform in environments where the subjects are well observable without (significant) attenuation or alteration, a dependable vision system must reckon with the entire spectrum of complex unconstrained and dynamic degraded outdoor environments. It is highly desirable to study to what extent, and in what sense, such challenging visual conditions can be coped with, for the goal of achieving robust visual sensing.

Our challenge is based on the A2I2-Haze, the first real haze dataset with in-situ smoke measurement aligned to aerial and ground imagery.

Semi-supervised Action Recognition from Dark Videos

Action recognition has been applied in various fields, yet most current action recognition methods focus on videos shot under normal illumination. In the meantime, there are various scenarios where videos shot under adverse illumination is unavoidable, such as night surveillance, and self-driving at night. It is therefore highly desirable to explore possible methods that are robust and could cope with these scenarios. Meanwhile, given the high cost of video annotation, it would be unlikely that such videos are labeled exhaustively in real-world scenarios. With the availability of large-scale labeled videos shot under normal illumination, this challenge aims to fully utilize these videos for constructing models robust to the dark environment.

This challenge is based on the ARID dataset, the first real-world dark videos for action recognition in dark environments.

Atmospheric Turbulence Mitigation

The theories of turbulence and propagation of light through random media have been studied for the better part of a century. Yet progress for associated image reconstruction algorithms has been slow, as the turbulence mitigation problem has not thoroughly been given the modern treatments of advanced image processing approaches (e.g., deep learning methods) that have positively impacted a wide variety of other imaging domains (e.g., classification).

This challenge aims to promote the development of new image reconstruction algorithms for incoherent imaging through anisoplanatic turbulence.