I have tested different methods for splitting the video in seperate streams and sequences of images. In the final experiments, presented here, I splitted the video export from 360Insta studio with the 360-2d tool in the VDSC Video Editor. The three video-files for left, frontal and right view were consequently converted in image sequences by sampling 1 frame per 2 seconds, resulting in an image set of 1477 images (5760x2488 pixels). The first tests were performed with Jawset Postshot. The extraction of keypoints and their alignment in a pointcloud took one or two hours computation time. The registration of the cameras took another 4 or 5 hours. The final step, the computation of the Gaussians splats for rendering could be followed real time by inspecting the updates of the iterative rendering process. I decided to stop this process after about half an hour for close inspection. Navigation in this model appeared to be very difficult because the scene was only recognizable from a limited number of view points. Continuation of this process did not yield better results. The best renders were obtained by choosing a viewpoint that is close to one of the camera's. The same image sets were also used as input for RealityCapture. I have used the fully automated workflow and after many hours of computation, I got a roughly detailed model with colored textures. Unfortunately, I could only view this model in the RealityCapture software. I couldn't find a way to get a managable export of this model. Therefore, I decided to show some screenshots in the following pictures.

Discussion and conclusions

• In this experiment, getting a point cloud and camara positions seemed to be no problem. The point cloud could be useful as a tool for navigation in the webviewer for 360 degrees video. This idea will be explored in further experiments.
• However, rendering Gaussian splats (Jawset Postshot) or textured meshes (RealityCapture) did appear to be a big problem. The results are not really surprising, because the glass panels, ligthing and foliage in this scene are well known as problematic for photogrammetry applications.
• Producers of 360 video cameras make an effort to produce a spherical image from planar images recorded by multiple sensors. In this experiment, these spherical images are converted to planar images. In all this processing, there is a risk that important image features will get lost or distorted. Better results might be obtained by methods that can avoid all these processing steps.