We provide a devkit to download, extract, and convert the challenge datasets into a unified format. This is done by first specifying the target root directory for all RVC datasets using an environment variable
export RVC_DATA_DIR=/path/to/rvc_dataroot
Now you can execute the download script download_obj_det.sh
which will download most of the RVC datasets.
You need to manually register and download the Mapillary Vistas (Research Edition) dataset: https://www.mapillary.com/dataset/vistas
You will receive an email with download instructions. Save/Move the downloaded zip file into the folder ${RVC_DATA_DIR}/mvd.
After successfully downloading all datasets, execute this script to extract and delete clean up files: extract_and_cleanup.sh
RVC does not force you to remap the datasets in a certain way. We do provide a "best-effort" mapping, which can be a good starting point. This mapping will contain overlapping classes and some dataset entries might miss relevant labels (as they were annotated using different policies/mixed hierarchical levels). Combine and remap datasets by executing the script
remap_obj_det.sh
The above step creates a joint training and a separate joint validation json file in COCO Object Detection format (only bbox entries, without "segmentation" entries):
http://cocodataset.org/#format-data
The "file_name" tag of each image entry has been prepended with the relative path calculated from RVC_DATA_DIR. These files can directly be used in your object detector training framework.
This repo will be updated as soon as the submission support is ready. See robustvision.net for news.