DrIFT: Autonomous Drone Dataset with Integrated Real and Synthetic Data, Flexible Views, and Transformed Domains

TL;DR

DrIFT tackles the problem of domain shifts in drone detection by introducing a 14-domain dataset with explicit background segmentation to study BG shift and to enable BG-wise evaluation. It introduces MCDO-map, a computationally efficient uncertainty metric derived from MC dropout, and an uncertainty-aware UDA framework that leverages this map for domain alignment, yielding improvements over state-of-the-art on DrIFT. The dataset design spans four DS axes—PoV, synthetic-to-real, season, and weather—facilitating systematic analysis and comparisons via BG-wise metrics and KL-divergence correlations. Empirical results demonstrate that MCDO-map tracks domain shift via KL divergence and that the proposed uncertainty-aware UDA enhances robustness across diverse DS scenarios, offering practical benefits for reliable drone detection in real-world, adverse conditions.

Abstract

Dependable visual drone detection is crucial for the secure integration of drones into the airspace. However, drone detection accuracy is significantly affected by domain shifts due to environmental changes, varied points of view, and background shifts. To address these challenges, we present the DrIFT dataset, specifically developed for visual drone detection under domain shifts. DrIFT includes fourteen distinct domains, each characterized by shifts in point of view, synthetic-to-real data, season, and adverse weather. DrIFT uniquely emphasizes background shift by providing background segmentation maps to enable background-wise metrics and evaluation. Our new uncertainty estimation metric, MCDO-map, features lower postprocessing complexity, surpassing traditional methods. We use the MCDO-map in our uncertainty-aware unsupervised domain adaptation method, demonstrating superior performance to SOTA unsupervised domain adaptation techniques. The dataset is available at: https://github.com/CARG-uOttawa/DrIFT.git.

Figures (10)

• Figure 1: Samples of the DrIFT dataset: a) aerial PoVs in different seasons for both real and synthetic data, b) ground PoV real and synthetic data that has been recorded in winter with a sky, tree, or ground background, c) adverse weather in a synthetic environment for both aerial and ground PoV data, and d) ground truth bounding boxes and background segmentation maps have been illustrated in first two rows. The utilized drones have been also depicted in the last row.
• Figure 2: Uncertainty-aware UDA framework: In addition to supervised learning on the source domain, the concatenated std and entropy maps are input into the discriminator as part of the adversarial learning process. Magnified regions around detections are shown for better visualization. Colorbars are placed on the right side of the std and entropy maps. Green-solid and red-dash lines represent the source and target domain paths, respectively. All detections from multiple iterations are displayed to illustrate the generation and behavior of the uncertainty maps.
• Figure 3: Correlation heatmap: The KL divergence of feature maps distributions and all metrics are calculated between the source (domain I in \ref{['tab:shiftresults']}) and target domains. MCDO-map's high positive correlation with KL divergence shows high capabilities of MCDO-map to capture the DS.
• Figure S1: Hierarchical sunburst chart of the DrIFT dataset: The DrIFT dataset contains aerial and ground views in real-world and simulated environments. There are numerous domains based on the various seasons and weather. The chart displays the number and percentage of the samples within the parent category. Adv: adverse
• Figure S2: Number of existing background samples in each domain: We aim to maintain an equal number of background samples for each domain's training and validation sets, unless the distribution of real data prevents us from adhering to this guideline.