Detecção de objetos de trânsito da simulação para o mundo real: um estudo de adaptação de domínio não supervisionada em nível de pixel e de características

Abstract

Training object detection models for real-world applications often requires large amounts of annotated data, which can be costly and time-consuming to obtain. Synthetic datasets provide a scalable alternative, enabling automatic annotation and controlled data generation. However, models trained exclusively on synthetic data typically suffer from performance degradation when applied to real-world images due to domain shift. This dissertation investigates simulation-to-real domain adaptation for object detection, focusing on the combination of pixel-level and feature-level adaptation strategies. At the pixel level, CycleGAN is used to translate synthetic images into the visual style of real-world datasets, reducing low-level discrepancies between domains. At the feature level, the ConfMix framework is adopted as a baseline for unsupervised domain adaptation, leveraging confidence-based mixing between labeled source samples and pseudo-labeled target samples. Building upon these approaches, this work proposes TransConfMix and TransConfMix-2Stage variants, which incorporate translated synthetic images into the adaptation process, and a KL-regularized extension (ConfMix-Distill) that enforces consistency between predictions obtained from original and translated versions of the same sample. These modifications aim to improve adaptation stability and enhance model generalization under domain shift. Experiments are conducted on multiple synthetic-to-real scenarios, with DOLPHINS and SIM10K as source domains and Cityscapes and nuScenes as target domains. The results demonstrate that feature-level adaptation consistently improves over the source-only baseline across the evaluated scenarios, and that pixel-level translation can further support this process when translation quality is sufficiently high. However, the proposed hybrid extensions did not consistently surpass the original ConfMix framework, suggesting that the interaction between pixel- and feature-level adaptation is highly sensitive to domain characteristics and translation fidelity. Overall, this work provides a systematic analysis of the interaction between pixel-level and feature-level domain adaptation, showing that their combined effect in object detection is strongly dependent on translation fidelity and domain characteristics.