Cross-Dataset Insights for Fine-Grained Vehicle Orientation Prediction

This dataset provides supplementary annotation files for cross-dataset fine-grained vehicle orientation estimation using the Car Full View dataset (CFV) and the Freiburg Static Cars 52 v1.1 dataset, also referred to as UnsupCar. The release contains:(1) re-annotated vehicle bounding box labels for the CFV dataset;(2) re-annotated vehicle bounding box labels for the Freiburg Static Cars 52 v1.1 (UnsupCar) dataset; and(3) harmonized vehicle orientation labels for the Freiburg Static Cars 52 v1.1 (UnsupCar) dataset. The bounding box annotations were generated using a YOLO-based detector and subsequently manually verified and corrected using LabelMe to improve consistency and accuracy across datasets. The harmonized orientation labels for UnsupCar were produced to align its angular annotation convention with the CFV label space. The harmonization procedure includes angular direction/reference correction and an orientation-dependent periodic bias correction, enabling fairer cross-dataset evaluation. This release does not include the original image datasets. Users should obtain the CFV (https://github.com/fort-cyber/car-orientation) and Freiburg Static Cars 52 v1.1 (https://github.com/lmb-freiburg/unsup-car-dataset) image data from their original sources and use the files in this record as supplementary annotations. The purpose of this dataset is to support reproducible benchmarking of vehicle orientation estimation, particularly under cross-dataset transfer settings where annotation-label incompatibility and crop-definition differences can otherwise confound model evaluation.