Análise de estratégias de active learning para detecção de objetos em cenários de trânsito

Abstract

Technological advancements have made it easier to acquire vast amounts of data, primarily used to train supervised learning algorithms. This makes it crucial to study techniques for selecting the most valuable information for model training and to identify the scenarios where such techniques can effectively improve performance. This work explores several active learning strategies for object detection, particularly in traffic scenarios, which present greater visual complexity due to diverse conditions such as varying luminosity, occlusions, and different seasons. To this end, a pipeline is proposed that integrates YOLO and LightlyOne to understand the behavior of uncertainty, diversity, and balancing strategies. The approach is tested on datasets with varying scenarios. The results demonstrate that using these strategies does not always guarantee improved model performance. The outcome is intrinsically linked to whether the primary challenge is the difficulty of object detection or class confusion—a characteristic inherent to the data context and the model being used.