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

Fardad Dadboud, Hamid Azad, Varun Mehta, Miodrag Bolic, Iraj ManteghWACV 2025

TL;DR

This paper introduces DrIFT, a novel dataset for drone detection that specifically addresses domain shifts caused by variations in viewpoint, synthetic-to-real data, seasons, and weather. They also propose a new uncertainty estimation metric, MCDO-map, and demonstrate its effectiveness in an uncertainty-aware unsupervised domain adaptation (UDA) method that outperforms state-of-the-art approaches.

What Problem Does This Paper Address?

The increasing use of drones in various applications necessitates reliable and accurate drone detection systems. However, these systems often struggle with "domain shifts" – changes in the environment or data characteristics between training and deployment. These shifts can arise from different camera viewpoints (PoV), variations between real and synthetic data, seasonal changes, and adverse weather conditions. Collecting and annotating data for every possible scenario is impractical and costly. While Unsupervised Domain Adaptation (UDA) offers a solution by transferring knowledge from labeled source domains to unlabeled target domains, existing datasets lack the diversity to systematically study these domain shifts, and current UDA methods can still be challenged by unforeseen situations.

Methods

The core contributions of this paper are a new dataset and a novel uncertainty estimation method integrated into a UDA framework.

  • DrIFT Dataset: This dataset is designed to isolate and study four key domain shift types: PoV, synthetic-to-real, season, and weather. It comprises fourteen distinct domains, offering a controlled environment for evaluating drone detection models under specific shifts. A key innovation is the inclusion of background segmentation maps, enabling background-wise metrics (e.g., AP for sky, tree, or ground backgrounds) to specifically analyze the impact of background shifts.
  • MCDO-map (Uncertainty Estimation): The authors propose a new uncertainty estimation method that leverages Monte Carlo Dropout (MCDO) but with a simplified pixel-wise score map approach. This method aims to capture epistemic uncertainty (due to data gaps or domain coverage) more efficiently than traditional MCDO techniques, reducing post-processing complexity.
  • Uncertainty-Aware UDA: This method integrates the MCDO-map into an adversarial UDA framework. The uncertainty maps are used to guide the adaptation process, focusing on regions where domain shifts have the most impact. The goal is to align feature distributions between source and target domains while minimizing the domain discriminator's ability to distinguish them.

Datasets

  • DrIFT Dataset: This is the primary contribution of the paper. It includes real and synthetic data, covering aerial and ground viewpoints, various seasons, and adverse weather conditions. It also provides ground truth bounding boxes and background segmentation maps.
  • Existing Datasets (for comparison): The paper references and compares DrIFT to several existing drone detection datasets (e.g., [11], [55], [26], [66], [48]), highlighting their limitations in terms of domain diversity and systematic study of shifts.

Key Insights and Ideas

  • Background Shift is a Distinct Challenge: The paper argues that background shift is a significant factor affecting drone detection and proposes a novel way to measure it using segmentation maps, enabling more granular analysis.
  • MCDO-map Captures Domain Shift Effectively: The proposed MCDO-map shows a strong positive correlation with KL divergence of feature maps, indicating its ability to capture domain shifts. This is crucial for guiding UDA methods.
  • Uncertainty-Aware UDA Outperforms: Integrating uncertainty estimation (specifically MCDO-map) into UDA significantly improves performance, especially in challenging target domains with substantial shifts.
  • Balanced Domain Study is Crucial: The DrIFT dataset's design, with balanced sample distribution across domains, allows for a more systematic and reliable study of individual domain shift impacts.

Key Results

  • The DrIFT dataset comprises fourteen distinct domains, enabling the study of four major domain shift types.
  • The proposed MCDO-map method demonstrates superior uncertainty estimation capabilities compared to MCDO-NMS and Grad-loss, particularly in capturing domain shifts.
  • The uncertainty-aware UDA method developed by the authors outperforms state-of-the-art UDA methods on the DrIFT dataset, achieving higher AP scores in various challenging domain shift scenarios. For instance, in a sky-to-tree shift scenario, their method achieved an AP of 10.7 for the tree background, a significant improvement over source-only performance.

Limitations and Future Work

  • Source Domain Performance Degradation: The authors acknowledge that their UDA method, like many others, can lead to a slight degradation in performance on the source domain.
  • Future Work: The authors aim to explore more advanced domain adaptation techniques that can minimize this source domain performance degradation, further enhancing the robustness and applicability of their approach.

Summary updated

BibTeX
@inproceedings{dadboud2025drift,
  author = {Fardad Dadboud and Hamid Azad and Varun Mehta and Miodrag Bolic and Iraj Mantegh},
  title = {{DrIFT: Autonomous Drone Dataset with Integrated Real and Synthetic Data Flexible Views and Transformed Domains}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure