A Baseline for Cross-Domain Fine-Grained Vehicle Classification in a Supervised Partially Zero-Shot Setting

Fine-grained vehicle classification is an important task particularly for security applications like searching for cars of suspects who abuse stolen license plates. However, data privacy and the large number of existing car models render it highly difficult to create a large up-to-date dataset for fine-grained vehicle classification with surveillance images. While a large number of images of vehicles are available in the web due to car selling sites, they have a perspective which is vastly different to surveillance images. Domain adaptation is the field of research that uses domain-wise inappropriate images for training of
classification models with the target of running accurate inference on images of a different domain. Since the widely considered unsupervised and semisupervised domain adaptation settings are unrealistic for fine-grained vehicle
classification, we establish a baseline for cross-domain fine-grained vehicle classification in a supervised partially zero-shot setting. Our results indicate that existing domain adaptation methods like domain adversarial training and triplet loss are still advantageous for this setting and we show the benefit of distance-based classification for this task.