Dronescapes Dataset Overview

Dataset Download

Pre‑processed Dataset Download

• Option 1: Download the pre‑processed dataset from the HuggingFace repository.

git lfs install
git clone https://huggingface.co/datasets/Meehai/dronescapes

Note: The dataset size is about 300 GB.

Generate Dataset from Raw Videos and Basic Labels

• Option 2: Recommended for understanding dataset creation or adding new videos and representations.
  • 1.2.1 Raw Videos: Provide 4K videos and pre‑processed 540p versions.
  • 1.2.2 Semantic Segmentation Labels: Manually annotated and then propagated using segprop.
  • 1.2.3 Generate Other Representations: Use video‑representations‑extractor to create additional labels.
  • 1.2.4 Convert to Mask2Former: Run scripts/convert_m2f_to_dronescapes.py to convert Mapillary or COCO classes to Dronescapes‑compatible 8 classes.
  • 1.2.5 Check Count Consistency: Execute bash scripts/count_npz.sh raw_data/npz_540p to verify data consistency.
  • 1.2.6 Split into Train, Validation, Semi‑Supervised, and Test Sets: Use scripts/symlinks_from_txt_list.py with text files to partition data.
  • 1.2.7 Convert Camera Normals to World Normals (optional): Provide camera rotation matrices for conversion.

Data Usage

Using the Provided Viewer

• Basic usage:

python scripts/dronescapes_viewer.py data/test_set_annotated_only/

Semantic Segmentation Evaluation

• Evaluate using scripts/evaluate_semantic_segmentation.py.

python scripts/evaluate_semantic_segmentation.py y_dir gt_dir -o results.csv --classes C1 C2 .. Cn [--class_weights W1 W2 ... Wn] [--scenes s1 s2 ... sm]

The above information outlines the download, generation, conversion, splitting, and usage procedures for the Dronescapes dataset.