Abstract:

This thesis investigates the problem of reconstructing tree skeletons for agricultural robotics, comparing single-view image-based (Image to 3D) and volumetric (3D to 3D) methods. Accurate 3D modeling is essential for robotic tasks like pruning and harvesting, where understanding the underlying branch structure is critical. Using a custom-generated dataset of synthetic trees, we train encoder-decoder models to infer 3D skeletal structures under conditions of occlusion. Our results show that 3D to 3D reconstruction significantly outperforms image-based methods in terms of F1-score and IoU, particularly in occluded environments. We further evaluate the effects of architectural choices and loss functions, finding that skip connections and weighted binary cross-entropy enhance performance. These findings offer practical insights for deploying robust 3D perception systems in agricultural robotics.

Committee:

Oliver Kroemer(advisor)

George Kantor(advisor)

Shubham Tulsiani

John Kim