Research map / Detection and segmentation / Tracking
BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces BuckTales, a UAV dataset of wild blackbuck lekking for multi-object tracking and simultaneous-view re-identification, with detection and tracking baselines (YOLOv8, ByteTrack, OC-SORT, BoT-SORT).
Problem
Automated long-term tracking and re-identification of wild animals from multiple simultaneously flying UAVs lacks public benchmark datasets with ground truth.
Approach
A multi-UAV dataset of wild blackbuck lek videos (5.4K, 80 m altitude) with MOT tracks and cross-drone Re-ID ground truth built via a ML-assisted annotation workflow (YOLOv8 + ByteTrack propagated IDs verified by annotators), plus benchmarks of several trackers.
Key results
- BoT-SORT with osnet_x1_0 Re-ID achieves the best HOTA of 0.5378 and 153 ID switches on BuckTales MOT, versus ByteTrack HOTA 0.4950 and OC-SORT HOTA 0.4203.
- YOLOv8m at image size 5472 reaches mAP 0.624 on the blackbuck detection set, the best detector reported.
- The dataset contains 1.2 million bounding box annotations over 680 tracks in 12 videos at 5.4K resolution, plus 730 Re-ID tracks from 11 drone pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BoT-SORT instead of ByteTrack | BuckTales MOT tracking · BuckTales MOT | HOTA: 0.4950 -> 0.5286 | Helped |
| Use BoT-SORT instead of OC-SORT | BuckTales MOT tracking · BuckTales MOT | HOTA: 0.4203 -> 0.5286 | Helped |
| Add appearance-based Re-ID to BoT-SORT (ablation) | BoT-SORT · BuckTales MOT | IDF1: 0.9214 -> 0.9451 | Helped |
| Add appearance-based Re-ID to BoT-SORT (ablation) | BoT-SORT · BuckTales MOT | HOTA: 0.5286 -> 0.5378 | Helped |
| Replace no-Re-ID BoT-SORT with osnet_x1_0 (ablation) | BoT-SORT · BuckTales MOT | IDsw: 171 -> 153 | Helped |
| Use osnet_x0_25 instead of no Re-ID (ablation) | BoT-SORT · BuckTales MOT | IDF1: 0.9214 -> 0.9418 | Helped |
| Use osnet_x0_25 instead of no Re-ID (ablation) | BoT-SORT · BuckTales MOT | HOTA: 0.5286 -> 0.5365 | Helped |
| Use osnet_ain_x1_0 instead of no Re-ID (ablation) | BoT-SORT · BuckTales MOT | IDF1: 0.9214 -> 0.9435 | Helped |
| Use lmbn_n instead of no Re-ID (ablation) | BoT-SORT · BuckTales MOT | HOTA: 0.5286 -> 0.5355 | Helped |
| Use osnet_x1_0 instead of osnet_ain_x1_0 (ablation) | BoT-SORT · BuckTales MOT | IDF1: 0.9435 -> 0.9451 | Helped |
| Use osnet_x1_0 instead of osnet_ain_x1_0 (ablation) | BoT-SORT · BuckTales MOT | IDsw: 150 -> 153 | Hurt |
| Train YOLOv8m at half image resolution (ablation) | YOLOv8m detection · BuckTales detection | mAP | No clear effect |
Surprising findings
- MOTA is high (above 0.93 for all trackers) despite large differences in HOTA, because MOTA overlooks association errors and permits many identity switches.
- Re-ID tracking uses person-trained Re-ID models, and no Re-ID baseline was provided for the cross-drone Re-ID problem.
Limitations
- Bounding boxes are not always tight and consistent, reducing localisation IoU and HOTA.
- Sub-adult males and females look similar, so some class misassignments occur; metrics handle only a single class.
- Number of unique individuals in the Re-ID dataset is unknown, so Re-ID is limited to within video segments.
Experimental setup
Datasets: BuckTales detection dataset (320 images, 18.4K annotations) BuckTales MOT dataset (12 videos, 1.2M boxes, 680 tracks) BuckTales Re-ID dataset (730 tracks, 11 video pairs)
Benchmarks: BuckTales MOT
Metrics: HOTA DetA AssA MOTA MOTP IDF1 IDsw mAP
Baselines: ByteTrack OC-SORT BoT-SORT Fast R-CNN (Detectron2) YOLOv8m
Builds on: ByteTrack OC-SORT BoT-SORT YOLOv8
Closest papers by meaning
- MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps (CVPR 2025 · similarity 0.89)
- The PanAf-FGBG Dataset: Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition (CVPR 2025 · similarity 0.88)
- BirdCollect: A Comprehensive Benchmark for Analyzing Dense Bird Flock Attributes (AAAI 2024 · similarity 0.87)
- WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife (NeurIPS 2026 · similarity 0.86)
- WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 2024 · similarity 0.86)
- Generating the Wild: Individual-Consistent Image-to-Video Generation for Wildlife (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.