MicMac vs Nerfstudio

MicMac is free, open-source desktop software for reconstructing 3D models from imagery. Developed at IGN and ENSG within the LASTIG lab, it runs on Linux, macOS and Windows under the CECILL-B license. It supports frame cameras, pushbroom sensors and scanned analogue images, along with camera calibration models and dense image matching for deformation studies. Outputs can include point clouds, 3D models and ortho-imagery in a desired coordinate system. Intermediate results in open formats let users work at different stages, and quality indicators include bundle adjustment residuals, covariance matrices, parameter sensitivity and correlation images. The C3DC command computes a point cloud from oriented images; AperiCloud creates a sparse cloud with camera positions and orientations for viewing in MeshLab, but does not support 64-bit PLY. MicMac is adapted to big datasets and was initially built for experts and researchers, with use also described in architecture, archaeology and forestry. Documentation, tutorials, a wiki and user forum are available. Building from source involves multiple dependencies, and the Mac guide cautions that a current revision may not compile or may be unstable on a given OS.

Nerfstudio is a free, open-source API and toolkit for creating, training, and testing Neural Radiance Fields through modular components. Included methods are Nerfacto, Instant-NGP, NeRF, Mip-NeRF, TensoRF, and Splatfacto, its Gaussian Splatting implementation. Command-line tools process data, train and evaluate models, render videos, view trained scenes, and export models. A web viewer shows training in real time, lets users interact with scenes, renders custom camera trajectories, and creates videos. Share URLs allow others to watch progress, but those links are publicly accessible and can add latency. Documented capture workflows use COLMAP, Polycam, or Record3D to turn phone video into a 3D render. Installation is available through pip, source, or Docker. Docker requires NVIDIA GPU drivers supporting CUDA 11.8, and the quickstart requires an NVIDIA video card with CUDA installed. The project supports Tensorboard and Wandb logging, profiling, and debugging. It runs across Linux, Windows, web, API, and self-hosted environments, with Windows described as less tested and more fragile.