This model is best viewed by selecting the first person mode in navigation setting and slow speed (eg 0 or 1) and navigating at the neighborhood of the gray squares. It is computed from videos taken by a helmet-held DIY multi-camera (four Gopro Hero 3 at 100Hz forming a 360 camera) and by biking 2.5km in a city. The gray squares in the streets indicate the camera poses of 5229 keyframes in the videos.
The main steps of the reconstruction method are in my publications: synchronization and self-calibration (3DV‘15, CVIU‘17), structure-from-motion (BMVC‘07, IVC‘09), surface reconstruction (CVIU‘18 improved by IC3D‘19), curves are matched (PAMI‘02) and enforced in the surface, texturing (IC3D‘20). The input videos are the BikeCity1/BC1 dataset mentioned in BMVC‘16 and CVIU‘17.
The original textures are divided by 3. The multi-camera model is central, global shutter, with frame-accurate synchronization.
Here is an image of the video input
Comments