Multi-view manhole detection, recognition, and 3D localisation

Automatic Discovery and Geotagging of Objects from Street View Imagery

“…This flexible technique allows us to incorporate the approximate knowledge about the camera-to-object distances into the decision process. Furthermore, it can be seen as a generalization to the previously proposed methods [10,12] as it allows the integration of geometry/visual-based matching as additional energy terms. The designed MRF model operates on an irregular grid that consists of all of the intersections of view rays generated by object instances detected at the object segmentation step.…”

“…Specifically, as reported in [11], the testing of GPS precision has demonstrated a 1.770 m 95% confidence interval for horizontal positioning achieved by SPS with a single-frequency receiver. One important contribution of this work is that our MRF-based geotagging procedure allows us to leverage monocular depth estimates to automatically resolve complex scenes containing multiple instances of identical objects, unlike previous works that rely on visual matching [10,12]. The proposed pipeline is modular which makes it possible to replace segmentation and depth modules with alternative techniques or pretrained solutions for particular object families.…”

“…It is assumed that objects are discovered through both imaging modalities. Together with [12], these methods rely on a simplifying locally flat terrain model in order to estimate object locations from single street view images. LiDAR and 3D point cloud data have been used alongside street view imagery to perform land use classification [18], 3D segmentation [19], and change detection [20].…”

“…Several methods have been developed to map particular types of objects from street-level imagery: manholes [12], telecom assets [10], road signs [29], and traffic lights [30,31]. These methods rely on position triangulation from individual camera views to geolocate the considered road furniture elements.…”

“…This highlights the necessity for non-standard triangulation such as our MRF-based technique to enable object mapping. In the considered road cluster around 35% of the pole-containing scenes cannot be mapped using methods similar to those proposed in [10,12,29] due to the presence of multiple visually identical telegraph poles, which renders visual matching-based triangulation inefficient.…”