MiLSD: Micro Line-Segment Detector for Resource-Constrained Devices

Line segment detection is a fundamental component in various computer vision applications, including visual SLAM, 3D reconstruction, and industrial inspection. While recent deep learning advancements have boosted accuracy, even the smallest existing models typically demand several megabytes of memory, exceeding the capacity of low-cost microcontrollers (MCUs). This research introduces MiLSD, a detector specifically engineered for MCU-level constraints, aiming to maximize accuracy within a sub-megabyte budget. The study systematically compares three output representations within a compact fully-convolutional backbone, finding that the proposed F-Clip center-with-length-and-angle formulation performs most effectively at small model sizes. The work also investigates quantization, showing that 8-bit quantization largely preserves full-precision performance, whereas 4-bit quantization leads to significant degradation, particularly in angle regression. With a one-megabyte activation budget and inference enhancements like sub-pixel decoding and a lightweight verifier, MiLSD improves sAP10 on the ShanghaiTech Wireframe dataset from 10.6 to 24.1, demonstrating a substantial leap in performance for embedded vision systems.