Learning from Synthetic Vehicles

In this paper, we release the Simulated Articulated VEhicles Dataset (SAVED) which contains images of synthetic vehicles with moveable vehicle parts. SAVED consists of images that are much more relevant for vehicle-related pattern-recognition tasks than other popular pretraining datasets such as ImageNet. Compared to a model initialized with ImageNet weights, we show that a model pretrained using SAVED leads to much better performance when recognizing vehicle parts and orientation directly from an image. We also find that a multi-task pretraining approach using fine-grained geometric signals available in SAVED leads to significant improvements in performance. By pretraining on SAVED instead of ImageNet, we reduce the error rate of one of the state of the art vehicle orientation estimators by 51.2% when tested on real images. We release SAVED and instructions on its usage here ¹ . ¹ https://taesoo-kim.github.io/