• OA-DG consists of OA-Mix for data augmentation and OA-Loss for reducing the domain gap.
• OA-Mix is a data augmentation methods that increases image diversity while preserving important semantic feature with multi-level transformations and object-aware mixing.
• OA-Loss is a novel contrastive loss functions that reduces the domain gap by training semantic relations of foreground and background instances from multi-domain.
• Extensive experiments on standard benchmarks (Cityscapes-C and Diverse Weather Dataset) show that OA-DG outperforms SOTA methods on unseen target domains.
• OA-DG can be generally applied to improve robustness regardless of the augmentation set and object detector architectures.

Multi-Level Transformations

Multi-level transformation enhances the domain diversity of the augmented image without damaging the object locations.

• An image is randomly divided into several regions such as foreground, background, and random box regions.
• Different transformations, such as color and spatial transformations, are randomly applied to each region.
• Spatial transformations are applied at the foreground level to preserve the location of the object.

Object-Aware Mixing

In object detection, each object in an image has different characteristics, such as size, location, and color distribution. Depending on these object-specific characteristics, some transformations can damage the semantic features of an object.
Object-aware mixing enhances the affinity of the augmented image, mitigating the negative effects of transformations.

Examples

The left and right images are the original and OA-Mixed images, respectively:

We evaluated the robustness of our method for common corruptions and various weather conditions in urban scenes. mPC is an evaluation metric of robustness against out-of-distribution (OOD).

Cityscapes-C

DWD