#Community Spotlight - Samuel McFadden🌍 We're constantly amazed by how our community uses Mapillary imagery to push the boundaries of what's possible. Amazing work by Samuel on taking 360° image sequences and creating a fully explorable 3D scene using the vid2scene platform. Check out the flythrough below!👇 #Mapillary #3DReconstruction #GaussianSplatting #SpatialComputing #GIS
Happy spring all! Lately, I’ve been wondering: what if street-level maps weren’t flat panoramas, but fully explorable 3D worlds? So instead of jumping between stitched images, you could move continuously through real-world scenes. I’ve been working on a pipeline that does that. As a first test, I converted a street-level 360 image sequence from a bike path near where I live into a full 3D Gaussian Splat scene, shown in the video below. The process: I pulled a sequence of equirectangular 360 images from Mapillary's API and generated the 3D scene using a custom build of the vid2scene platform adapted for street reconstructions. The pipeline is still very delicate and requires some heavy image preprocessing to keep the reconstruction from exploding. It's not perfect yet (the details are a little too fuzzy and there's some heavy visual artifacts), but it works. And it raises an interesting question: Mapillary has billions of street-level images covering roads all over the world. Would it be possible to process the entire dataset into an open 3D world model? Is this the next phase of Google Maps? If you're working on 3D reconstruction, autonomous vehicle training data, urban planning, or anything that could benefit from photorealistic explorable 3D streets, I'd love to hear your thoughts. Link to the 3D scene is in the comments. And as always, credit to the PlayCanvas team for their SuperSplat editor which I used to generate the flythrough render. #GaussianSplatting #3DGS #3DReconstruction #SpatialComputing #StreetView #Mapillary