C2D2: Matching UAV images with cycle-consistency and channel-separation for joint detector and descriptor

Accurate feature matching in Unmanned Aerial Vehicle (UAV) imagery is fundamental to earth observation applications. Although jointly learning keypoint detectors and descriptors has become standard practice, existing methods face inherent challenges: keypoint repeatability and descriptor reliability optimize for different visual properties, while patch-based training losses impose locality constraints that limit global geometric consistency. This paper presents C2D2, a network that addresses these limitations through two key innovations. First, a Channel Separation Module (CSM) reduces cross-task interference by enabling task-specific channel emphasis. Second, a cycle consistency loss enforces global geometric constraints beyond local patches, enhancing descriptor robustness with minimal additional training overhead. Extensive evaluation on public benchmarks and a challenging UAV dataset demonstrates improvements on most evaluated metrics: 4.4% higher detector repeatability on HPatches and up to 7.3% improvement in localization recall on the Aachen Day–Night benchmark. These results , achieved with only 0.49M parameters, establish C2D2 as an efficient solution within the sparse feature learning family for demanding remote sensing applications.