Experiments with Circle Loss on AIC 2021’s Vehicle Retrieval Dataset
Info
Usage
Extract metadata
The dataset provided metadata in the form of an XML file train_label.xml which can be hard to processed. We first convert this into a more accessible JSON file.
The result will be saved as list/train_image_metadata.json.
Split data
Since we use the data above for training, evaluation, and testing, we split it into corresponding CSV files.
The results are stored in the list folder as CSVs file of tuples of (image_id, vehicle_id, cam_id):
reid_train.csv: contains the training datareid_query_[val/test]: contains the queries for evaluationreid_gallery_[val/test]: contains the gallery for evaluation