The ForestryAI at NIBIO team just dropped a game-changer! 💥 💣
SmartForest introduces ForestFormer3D – our new state-of-the-art transformer model for forest panoptic segmentation! It tackles semantic (ground, wood, leaf) and instance-level (individual trees!) segmentation in 3D forests like never before.
The pre-print is available here: https://arxiv.org/abs/2506.16991
You can access the project page here: https://bxiang233.github.io/FF3D/
Code and data will be published soon!
🔍Highlights:
✨ FOR-instance V2 – a next-level benchmark for complex, layered, and overlapping tree canopies 🌳🌳🌳
🧠 Transformer architecture – more complex architecture = improved performance
🎯 Instance + semantic-aware prompting – intelligent prompts = better results
🧩 other innovations – score-based block merging + one-to-many association
🏆 Beats the best – outperforms ForAINet, SegmentAnyTree, and TreeLearn 📊
🌲 With such advances, single tree forest management is getting closer and closer to becoming operational reality. ✨ A world of opportunities lies ahead.
Congratulations to Binbin Xiang for leading this groundbreaking study, and shoutout to our awesome team at NIBIO (Maciej Wielgosz, Stefano Puliti, Rasmus Astrup), and Czech collaborators Kamil Kral, Martin Krucek, and Azim Missarov from Silva Tarouca Research Institute for Landscape and Ornamental Gardening, Prague, Czech Republic.