FNI V2.0 for dronescapes: Authority (A:46), Popularity (P:58), Recency (R:24), Quality (Q:50). Semantic (S) is a query-time baseline scored live at search.
Visit the official website for more information: link. This dataset was introduced in our ICCV 2023 workshop paper: link. For citing, see at the end of the page.
Note: An fully-automated extended variant of this dataset (generating new modalities as inputs) is available at this repository: link.
1. Downloading the data
text
git lfs install # Make sure you have git-lfs installed (https://git-lfs.com)
git clone https://huggingface.co/datasets/Meehai/dronescapes
Note: the dataset has about 200GB, so it may take a while to clone it.
1.2 Low level data for the dataset (GPS, camera rotation matrices)
Click to expand
1.2.1 Convert Camera Normals to World Normals
This is an optional step, but for some use cases, it may be better to use world normals instead of camera normals, which
are provided by default in normals_sfm_manual202204. To convert, we provide camera rotation matrices in
raw_data/camera_matrics.tar.gz for all 8 scenes that also have SfM.
In order to convert, use this function (for each npz file):
The GPS location (lat/long/height) is in the raw_data/raw_camera_info.tar.gz. Each file over there is an archive over the
original videos, so the train/val/test splits files from raw_data/txt_files must be used to map