Event

Any3R Open-Sourced: A Single Set of Weights Uniformly Reconstructs Arbitrary Central Cameras

1 reports · 1 sources

Jingong Chen and co-authors from South Dakota State University and other institutions propose Any3R, which uses a dense ray field to describe arbitrary central cameras; a single forward pass can reconstruct a 3D point cloud from mixed-camera images and estimate poses. The inference code, 13 camera models, and model weights have been open-sourced. The method replaces the row and column indices of rotary position encoding with angular coordinates and weights the pixel loss by solid angle, enabling a single set of weights to cover pinhole, fisheye cameras with over 180° field of view, and 360° panoramic cameras. The authors report that on benchmarks covering 13 central camera types and 3 datasets, its reconstruction accuracy and pose accuracy both lead. The model has about 1.15 billion parameters, is fine-tuned from VGGT-Omega, and was trained for 40 epochs on 32 GH200 GPUs in about 27 hours; on a single RTX 5090, inference on 32 512×512 views takes about 1.1 seconds, and on 128 views about 9 seconds with a peak GPU memory of 17.7 GB. The open-source release also includes camera models that support reading COLMAP and OpenCV calibration, a command-line interface and Gradio demo, and benchmark evaluation code.