Adapting Synthetic Eyes: A Study of Pixel and Feature-Level UDA for Traffic Object Detection

• André N. Medeiros UFSCar
• Renato M. Silva USP
• Jurandy Almeida UFSCar
• Tiago A. Almeida UFSCar

Resumo

Training robust object detectors for autonomous driving requires vast amounts of annotated data, making the use of synthetic datasets an attractive alternative. However, models trained on synthetic data suffer from a significant performance drop when deployed in the real world due to the “sim-to-real” domain gap. Unsupervised Domain Adaptation (UDA) aims to solve this problem without requiring expensive target domain annotations. This paper conducts an empirical study comparing two leading but philosophically different UDA paradigms: pixel-level adaptation via image-to-image translation (CycleGAN) and feature-level adaptation via confidence-aware data mixing (ConfMix). We evaluate these methods on challenging synthetic-to-real adaptation tasks, using DOLPHINS and SIM10K as source domains, and Cityscapes and nuScenes as target domains. Our findings demonstrate that the feature-level mixing strategy of ConfMix provides more significant and robust performance gains than pixel-level translation with CycleGAN. Furthermore, we introduce and evaluate a hybrid method, TransConfMix, which yields mixed results, highlighting the complexities of combining these techniques. Our work provides clear evidence and practical guidance on the effectiveness of different UDA strategies, concluding that directly adapting the model's learning process is a more potent approach than preprocessing the data for this critical application.

Palavras-chave: Training, Adaptation models, Translation, Annotations, Vehicle detection, Detectors, Object detection, Feature extraction, Data models, Synthetic data