The BAMBI Dataset: Multimodal Nadir UAV-Recordings of Forest Wildlife (Part 2)

This dataset is part of the BAMBI (Biodiversity Airborne Monitoring Based on Intelligent UAV sampling; https://www.bambi.eco/) project. It contains multimodal nadir UAV recordings of forest wildlife captured with thermal and RGB sensors. The recordings where processed in terms of (1) undistorting the view, (2) removing FCC calibration phases of the thermal camera and (3) syncing the thermal and RGB videos based on the flights logs (still contains some temporal offsets). Each flight is stored as an individual ZIP archive named by its numeric flight ID (e.g. flight_0.zip, flight_1.zip, …). Inside each ZIP are the associated files for that flight (e.g. 0_mask_t.png, 0_mask_w.png, 0_matched_poses.json, 0_matched_processed.mp4, …). Due to the size of the full dataset it has been split into multiple parts. Please download all parts to obtain the complete dataset. You can find scripts for working with the dataset at https://github.com/bambi-eco/Dataset. Version 2 changes. Refined multi-object tracking labels. The ground-truth labels were re-reviewed: bounding boxes were corrected, additional key frames and tracks were added, and a number of flights received labels for the first time. Reviewed sex and occlusion labels. For a red deer subset, the gender and visibility columns were replaced by the reviewed labels of the BAMBI Matched Dataset (10.5281/zenodo.21291168, 10.5281/zenodo.21298044). The reviewed classes are male and female_juvenile; the latter is a single indivisible survey class covering females and calves and is recorded as gender=1 (Female). The age column is unchanged by this step and must not be read as a consequence of it. visibility is derived from the reviewed thermal occlusion ranges; frames outside a reviewed range keep their previous value. Species labels are unchanged. Where the two annotation efforts disagree about species, this dataset's own species labels stand. Track IDs are stable. Every track that existed in version 1 keeps its version 1 track_id; tracks added in version 2 use IDs above the highest ID version 1 used. Flight archives whose labels did not change are byte-identical to version 1.