DeepReefMap

Hugues Sibille (EPFL) and Jonathan Sauder (MIT and EPFL) maintain DeepReefMap under EPFL's ECEO group.¹ The method, for rapid 3D semantic mapping of reefs, was published in Methods in Ecology and Evolution in 2024.¹ From one video it produces a semantic 3D point cloud, an orthomosaic, per-class benthic cover statistics and an interactive viewer.¹ A semantic point cloud is a set of 3D points, each labelled with what it shows. An orthomosaic is a single map-like image stitched from many frames. Benthic cover is the share of the seabed each class occupies.

The pipeline rectifies frames with a camera profile, segments each frame (labels every pixel by class), reconstructs depth and camera pose, then exports the products.¹ It ships with profiles for the GoPro Hero 10 and 12.¹ Other cameras are calibrated with a built-in tool based on COLMAP, an open-source photogrammetry program.¹ Several reconstruction backends can be chosen. A basic one runs on CPU, higher-quality ones need a GPU (NVIDIA, AMD or, for some, Apple Silicon), and one has gated weights limited to non-commercial research use.¹ Segmentation defaults to a model trained on CoralscapesV2, a panoptic reef dataset of 2,433 annotated images from 55 Red Sea sites with 95 classes, released under Apache 2.0.^1,2 A panoptic dataset labels both individual objects and background regions.

DeepReefMap is a command-line tool in Python 3.10 to 3.12, licensed Apache 2.0.¹ A GPU is strongly recommended, and a separate desktop app exists.¹ Bundled third-party components and downloaded checkpoints have their own terms.¹ The project is public and actively developed, with workshop papers cited for 2025 and 2026, and its latest tagged release, v1.2.0, came out in October 2026.^1,3