After the success of SegmentAnyTree, we are excited to share SegmentAnyTreeV2 — the latest contribution by SmartForest to build robust, scalable, and operational AI tools for single tree forest inventory and management.
read the preprint 📰 : https://arxiv.org/abs/2606.08206
Built to work across sensors, platforms, and forest types, SegmentAnyTreeV2 expands from well-managed conifer forests to a much broader range of forest conditions, including structurally complex broadleaved, mixed, and tropical forests.
The model combines a PointTransformerV3 backbone with a tree-focused instance decoder, enabling stronger separation of individual trees even in dense and complex stands.
Key highlights:
📊 Introducing FOR-instance V3 a new and more CHALLENGING benchmark dataset that standardizes several open benchmarks
✅ SegmentAnyTreeV2 achieves SOTA performance –> more trees are correctly segmented
✅ Improved coverage –> better tree masks
✅ Improved semantic –> better leaf/wood separation
✅ Strong zero-shot generalization across unseen forest sites
Congratulations to Maciej Wielgosz, Stefano Puliti, Rasmus Astrup and a huge thanks to everyone involved in pushing 3D forest computer vision with open data and open models!