Cross-Domain Object Detection Using Unsupervised Image Translation

TL;DR

This work tackles unsupervised domain adaptation for object detection by proposing a two-stage framework that first learns unsupervised image-to-image translation to produce target-like images from annotated source data, and then trains a detector on this artificial dataset. By leveraging CycleGAN and AST-AdaIN, the method generates a fake-target dataset with inherited annotations, enabling training of a standard detector (Faster R-CNN) on target-domain data without target annotations. The results across synthetic-to-real, adverse weather, and cross-camera scenarios show the approach often outperforms state-of-the-art DA methods and narrows the gap to the upper bound achievable with target data, while offering greater interpretability due to pixel-level translation and decoupled training stages. This simple, interpretable strategy reduces implementation complexity and computational demands relative to end-to-end domain-adaptation models, with practical implications for deploying robust detectors in autonomous driving across diverse environments.

Abstract

Unsupervised domain adaptation for object detection addresses the adaption of detectors trained in a source domain to work accurately in an unseen target domain. Recently, methods approaching the alignment of the intermediate features proven to be promising, achieving state-of-the-art results. However, these methods are laborious to implement and hard to interpret. Although promising, there is still room for improvements to close the performance gap toward the upper-bound (when training with the target data). In this work, we propose a method to generate an artificial dataset in the target domain to train an object detector. We employed two unsupervised image translators (CycleGAN and an AdaIN-based model) using only annotated data from the source domain and non-annotated data from the target domain. Our key contributions are the proposal of a less complex yet more effective method that also has an improved interpretability. Results on real-world scenarios for autonomous driving show significant improvements, outperforming state-of-the-art methods in most cases, further closing the gap toward the upper-bound.

Figures (5)

• Figure 1: Overview of the proposed method. Firstly, an unsupervised image translation model is trained with unpaired source and target images. Then, the source image set is translated to its fake-target version. The annotations of source images are directly transferred to the fake-target images, composing the fake-target dataset. Finally, an object detector is trained resulting in an object detector trained on a domain without previous annotations.
• Figure 2: Samples of each dataset. The datasets samples are presented row-by-row in the following order (top to bottom): Cityscapes, Foggy Cityscapes, Sim10k and KITTI. The original aspect ratio of the images was preserved.
• Figure 3: Samples of translated images with their respective bounding boxes. The real source training images are shown in the first row and their respective fake-target versions are shown in the second and third row for the CycleGAN and AST-AdaIN models, respectively.
• Figure 4: Samples of translated images with their respective bounding boxes. The real source training images are shown in the first row and their respective fake-target versions are shown in the second and third row for the CycleGAN and AST-AdaIN models, respectively.
• Figure 5: Samples of translated images with their respective bounding boxes. Top: the translation from Cityscapes to KITTI, and vice-versa for the bottom one. The real source training images are shown in the first row and their respective fake-target versions are shown in the second and third row for the CycleGAN and AST-AdaIN models, respectively.