| dc.description.abstract |
In the field of autonomous vehicles (AV), it is crucial for the perceptual systems of the AVs to learn inter-domain adaptations in the absence of paired examples for detecting vehicular instances in unstructured real-world scenarios. One straightforward approach is to train the models directly on labeled synthetic datasets. However, this approach usually fails to achieve generality, owing to the domain bias between the real and fake databases of images. We therefore, propose a novel architecture that produces synthetic images based on cycle consistency, in the absence of labeled pair of images. We test the object detectors Detectron and SSD on four types of curated benchmark datasets to evaluate their robustness in detecting objects such as cars, bikes, and pedestrians on road. The four benchmark datasets contain a diversified set of image corruptions and a few variations are built using the proposed framework such as weather variations. The four datasets are PASCAL VOC and SYNTHIA dataset images, weather-translated images, variation-augmented images, and stylized renderings using Adain style transfer. After conducting extensive investigations, we observe a decreased classification loss when exposed to variable image quality. We also witness that augmenting the training datasets with variations in the wild aided in boosting the generalizing capability of the object detectors. The boost in performance is testified by the testing results showing better mAP. |
en_US |