One photo in, a 3D tree out: tree masks, a radial bounding volume, and the reconstructed model. Figure: Li et al., “Learning to Reconstruct Botanical Trees from Single Images”, ACM TOG 2021.
Working Trees
Turning photographs into 3D trees
Farmers and ranchers can get paid for the carbon their trees store, if they can prove how much there is. Working Trees, a Stanford StartX company, lets them measure it with a phone, and I built the pipeline that turns a photo into a 3D tree.
It started from published research. I worked with professors and grad students at Purdue to adapt their method for reconstructing a tree from a single photo, and fine-tuned SAM and Depth Anything on our trees with reinforcement learning. For video, I built reconstruction on NeRFs and Gaussian splats that reached 95% scale accuracy.
The US beta went well. Then we launched in Brazil, where many of our farmers had cheap phones with weak cameras, often on patchy signal, and a pipeline built for none of that struggled. After regular calls with the Brazil team to see how they actually worked, I rearchitected the models to run on low-end phones and halved inference time.