DRIFT Iceberg Tracking Dataset: Time-Lapse Images, Annotations, and Models

Time-lapse images, sparse iceberg track annotations, and trained models used in the DRIFT paper (Jaeger-Kufel et al., ISPRS Open Journal of Photogrammetry and Remote Sensing, under review). The dataset contains 18 sequences of 50 images at a 2-minute interval from Eqalorutsit Kangilliit Sermiat (EKaS, two camera views, 2023 and 2024) and Eqip Sermia (2015), covering clear, moderate, ice mélange, twilight, and night conditions. It includes 45,000 annotated boxes of 972 iceberg tracks in MOTChallenge format, the trained Faster R-CNN detector and DINOv2 appearance encoder, and the detections, embeddings, and DRIFT tracks of all sequences. See README.md for the structure and format. Code: https://github.com/marco-jaeger/drift