PhotosVideosAllPeople
A video with on-screen text. Six synchronized lab cameras film a chickadee wearing a neural-recording headstage, and the views merge into BEAST3D's moving 3D reconstruction of the bird. A title card reads "BEAST3D: Animal behavioral analysis and neural encoding from multi-view video via Gaussian splatting", NeurIPS 2026 Oral. General 3D models (VGGT, E-RayZer) give noisy point clouds full of background, while BEAST3D's show only the animal. On a head-fixed mouse filmed by six cameras, one camera is hidden, BEAST3D predicts 3D Gaussians from the other five, and its render from the hidden camera plays next to the real video. 3D renders of a mouse, rat, chickadee and human follow. A slide asks "What can you do with BEAST3D?" and lists three downstream tasks: results for novel views and pose estimation, then a moving 3D point cloud of the bird next to recorded and predicted activity of 52 hippocampal neurons, and a bar chart of bits per spike. It ends with the title, authors and links.
Video, no description
Video, no description
Three reconstructions of the same scan with 1% noise: the noisy input, Instant-NGP with the noise baked into the surface, and M-plicits with a clean surface.