Fascinating work and inspiring application of the underlying DINOv3 image segmentation model!
The blog post and paper [1] describe a promising approach to solving related problems at previously impossible scale and quality: I am currently exploring methods to better represent seasonal land cover changes that would improve wind power generation forecasting and this paper provides a great starting point.
I hope DINOv3 can inspire more work like this - and I would encourage any curious mind to play with that model! I was amazed by its capability to distinguish between fine object details. For example, in a photo of a bicycle, the patch embeddings cleanly separated the background from the individual spokes of the wheel.
I gave a talk about the paper in our internal journal club recently (we work on similar problems, usually using stereo imagery though).
It's a nice piece of work. I especially like the sections on data cleaning and registration, as that seemed to have been one of the limiting factors of the previous approaches.
I am sceptical about how accurately you can predict heights for specific trees from mono-images, but I think for cases where you just need to be right on average (e.g. biomass estimation, fuel load estimates) it's a great approach.
I think they were buying carbon offsets at some point and trying to validate that the countries and organizations that were selling the carbon offset were not cutting down those trees, effectively profiting twice.
14 comments
The blog post and paper [1] describe a promising approach to solving related problems at previously impossible scale and quality: I am currently exploring methods to better represent seasonal land cover changes that would improve wind power generation forecasting and this paper provides a great starting point.
I hope DINOv3 can inspire more work like this - and I would encourage any curious mind to play with that model! I was amazed by its capability to distinguish between fine object details. For example, in a photo of a bicycle, the patch embeddings cleanly separated the background from the individual spokes of the wheel.
[1] https://arxiv.org/abs/2603.06382
It's a nice piece of work. I especially like the sections on data cleaning and registration, as that seemed to have been one of the limiting factors of the previous approaches.
I am sceptical about how accurately you can predict heights for specific trees from mono-images, but I think for cases where you just need to be right on average (e.g. biomass estimation, fuel load estimates) it's a great approach.