Indoor Defect Region Identification Using an Omnidirectional Camera and Building Information Modeling

An omnidirectional (i.e., 360°) camera is an efficient device that can capture the status of a room with a single shot. To detect objects in a spherical image captured by a 360° camera, the image should be flattened and divided into patches reflecting normal fields of view (NFoV). However, detecting indoor defects in omnidirectional camera images is difficult because they are relatively small and span multiple patches. Another challenge is to set the appropriate size for an NFoV patch. To overcome these challenges, this paper proposes a method to locate possible regions of indoor defects using building information modeling (BIM). The core idea is to subtract a 360° camera image from a photorealistically rendered BIM model image of the same location. Bounding boxes are generated around the areas where differences are detected. The proposed method was tested in a single room with artificially implanted cracks. In the experiments, two different omnidirectional cameras were used. The image classification algorithm was trained on open crack datasets. The results showed that the proposed method improved the F1-score from 0.15 to 0.39 and recall from 0.16 to 0.87. The proposed method could detect more cracks while reducing the number of patches needed for indoor crack inspection compared to the traditional method.