MCDet: A motion-centric spatiotemporal framework for dim moving object detection

Detecting dim moving objects is fundamentally challenging because their appearance is extremely weak in individual frames and easily overwhelmed by noise and background clutter. Existing methods, including single-frame detectors and multi-frame approaches, mainly rely on the assumption that reliable spatial features can be extracted per frame. This assumption often fails under low signal-to-noise conditions. In this work, we adopt a motion-centric perspective and treat motion as the primary object of interest rather than an auxiliary cue. Based on this idea, we propose MCDet, a unified spatiotemporal detection framework. MCDet models targets as trajectory volumes in a 3D space–time domain and performs dense regression directly on spatiotemporal features. The model jointly encodes spatial and temporal information from early stages. It further uses a decoupled prediction head to separate global trajectory modeling from frame-wise localization. Extensive experiments on four public benchmarks and real-world astronomical imagery demonstrate that MCDet consistently outperforms state-of-the-art methods. It achieves higher accuracy and produces fewer false alarms. Code and datasets will be made publicly available at https://github.com/CastleChen339/MCDet.