Facebook Research releases MapAnything, 3D reconstruction from images (github.com)

Meta (Facebook) Research released MapAnything, an end-to-end transformer that directly regresses factored metric 3D scene geometry from images and optional geometric inputs. Unlike multi-stage SfM/MVS pipelines, a single feed‑forward MapAnything model handles over a dozen tasks (multi-image SfM, multi-view stereo, monocular metric depth, registration, depth completion, etc.) and accepts flexible inputs—RGB, intrinsics or ray directions, depth maps, and camera poses—to produce per-pixel 3D points, Z-depth, depth-along-ray, recovered intrinsics and poses, confidence masks, and a learned metric scaling factor. The code, pretrained models (including an Apache 2.0 option), demos (Hugging Face/Gradio/Rerun), and COLMAP/GSplat export tools are available on GitHub, with mixed-precision and memory-efficient inference modes (bf16/fp16; supports large view counts up to thousands given sufficient memory). This matters because it unifies many reconstruction subtasks into a single, trainable model that can recover metric-scale geometry and camera parameters directly, simplifying pipelines for mapping, robotics, AR, and photogrammetry. Key technical details include strict input conventions (OpenCV cam2world poses, depth requires intrinsics or ray directions), per-pixel confidence and non-ambiguity masks, and an API that enforces input constraints while allowing 64 underlying input configurations. The release provides practical tooling for benchmarking and visualization, but users should note nontrivial compute/memory needs for high‑view reconstructions and the usual dependency on diverse training data for robust metric accuracy.