Dronescapes Experts dataset

This dataset is an extension of the original dronescapes dataset with new modalities generated using VRE 100% from scratch (aka pretrained experts). The only data that is not generable by VRE is the Ground Truth: semantic (human annotated), depth & normals (SfM) that is inherited from the original dataset for evaluation purposes only.

1. Downloading the data

Option 1. Download the pre-processed dataset from HuggingFace repository

git lfs install # Make sure you have git-lfs installed (https://git-lfs.com)
git clone https://huggingface.co/datasets/Meehai/dronescapes

Option 2. Generate all the modalities from raw videos

Follow the instructions under this file.

Note: you can generate all the data except semantic_segprop8 (human annotated), depth_sfm_manual202204 and
normals_sfm_manual202204 (SfM tool was used).

2. Using the data

As per the split from the paper:

<img src="split.png", width="500px">

The data is in data/* (if you used git clone) (it should match even if you download from huggingface).

2.1 Using the provided viewer

The simplest way to explore the data is to use the provided notebook. Upon running
it, you should get a collage with all the default tasks, like the picture at the top.

For a CLI-only method, you can use the VRE reader as well:

vre_reader data/test_set_annotated_only/ --config_path vre_dronescapes/cfg.yaml -I vre_dronescapes/semantic_mapper.py:get_new_semantic_mapped_tasks

3. Evaluation

See the original dronescapes evaluation description & benchmark for this.