ACDC: The Adverse Conditions Dataset With Correspondences for Robust Semantic Driving Scene Perception.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Level-5 driving automation necessitates advanced visual perception capable of handling diverse environmental conditions.
• Existing datasets for semantic perception in driving are limited by normal-condition bias or small scale.
• There is a critical need for comprehensive datasets that capture adverse visual scenarios.

Purpose of the Study:

• Introduce the Adverse Conditions Dataset with Correspondences (ACDC) for robust visual perception in autonomous driving.
• Facilitate training and evaluation of semantic perception models under adverse conditions.
• Enable research into uncertainty-aware semantic segmentation.

Main Methods:

• Compiled ACDC, a dataset of 8012 images, with 4006 images equally distributed across fog, nighttime, rain, and snow conditions.
• Provided pixel-level panoptic annotations for adverse-condition images and corresponding normal-condition images.
• Included binary masks for intra-image uncertainty assessment and supported standard and novel segmentation tasks.

Main Results:

• ACDC presents significant challenges to current supervised and unsupervised state-of-the-art methods.
• Empirical studies highlight the dataset's utility in identifying weaknesses in existing approaches.
• The dataset's diverse adverse conditions reveal performance gaps in semantic perception models.

Conclusions:

• ACDC is a valuable resource for advancing visual perception in autonomous driving under adverse conditions.
• The dataset will guide future research and development of more resilient perception systems.
• Public availability of ACDC and its benchmark will accelerate progress in the field.