🌲 ForestMamba on 3Dtrees.earth Platform

I am happy to share that ForestMamba has been integrated into 3Dtrees.earth, an open platform for sharing and processing terrestrial LiDAR (TLS) scans of forests.

ForestMamba now runs as part of the platform’s processing pipeline, performing semantic leaf/wood separation and single-tree instance segmentation on the uploaded forest point clouds. The model is applied automatically to newly uploaded scans and is being applied retroactively to the existing archive of nearly 2,000 datasets, together with a per-point segmentation confidence.

🌿 Why Mamba for forest point clouds?

Forest TLS scans are large, dense, and highly occluded, which makes attention-based models expensive to run at full scan scale. ForestMamba combines geometry-guided queries built from forest-specific structural cues with a sparse Mamba backbone, giving linear-time complexity and substantially lower memory use than Transformer-based alternatives — which is exactly what makes processing at platform scale practical.

More details are in our preprint: ForestMamba: Sparse Mamba with Geometry-guided Queries for 3D Forest Point Cloud Segmentation.

🎥 Announcement

Seeing a research prototype move into a platform that the forest-monitoring community actually uses every day is very rewarding. Many thanks to the 3Dtrees.earth team and to all my collaborators on this work — and if you try it on your own scans, I would love to hear how it performs!