ULSD: Unified line segment detection across pinhole, fisheye, and spherical cameras

EVOLIN Benchmark: Evaluation of Line Detection and Association

“…Different architectures have been proposed for line detection, such as transformer-based [27], fully convolutional [31], and having a trainable module that performs the Hough transform [24]. Various line representations have also been proposed, such as tri-point [25] and Bezier curves [29]. In addition, real-time detectors have been proposed [28], [30].…”

“…Line Detectors. We have compared and evaluated a total of 17 algorithms: AFM [21], ELSED [19], F-Clip [31], HAWP [26], HT-LCNN [24], L-CNN [22], LETR [27], LSDNet [30], LSWMS [74], M-LSD [28], TP-LSD [25], ULSD [29], FSG [18], and SOLD2 [61]. From OpenCV, the implementations for the algorithms: LSD [13], EDLines [15], and FLD [16].…”

LF-VIO: A Visual-Inertial-Odometry Framework for Large Field-of-View Cameras with Negative Plane

“…Pautrat et al proposed SOLD2 [19], a learning-based method that jointly detects and describes line segments in an image in a single network without the need of annotating line labels. Li et al [11] proposed a Unified Line Segment Detection (ULSD), which aims to detect line segments in both distorted and undistorted images.…”

“…Localization error: The localization error with tolerance ϵ refers to the average distance between a curve segment and its re-detection in the second image, considering only the matched curve segments. Comparison on the public datasets: Targeting omnidirectional data, we compare our proposed OCSD method against the state-of-the-art ULSD [11], which is a learning-based curve detection method supporting different camera models. The results are compared on the SUN360 indoor dataset [29] and the CVRG-Pano outdoor dataset [30] respectively, as shown in Table I and Table II.…”

“…Line extraction is a critical step in point-line odometry, which is a technique used to estimate the motion of a camera or robot by tracking the 3D positions of feature points and lines over time [11], [12]. In this technique, line features are extracted from the images captured by the camera, and they are matched with the corresponding lines in the next image frame to compute the camera's motion.…”