My Bachelor's Thesis Project

There are two parts to this project:

1. Evaluating three different surface normal predictors: Three Filters To Normal, Surface Normal Uncertainty, and a module I made inspired by the depth-to-normal refinement in GeoNet

And here is my full paper for more details: Lizzy.pdf

The paper is mostly finished, save for the introduction; I will update it once the final version is done. The final version of the paper was uploaded (2024-09-11)

The error metrics I used are the following:

$$ \text{Angle error} = \varphi = \arccos{\frac{y \cdot \hat{y}}{|y| |\hat{y}|} } $$

$$ \text{Mean error} = \frac{1}{n} \sum \varphi $$

$$ \text{Median error} = \text{Median}(\varphi) $$

$$ \text{RMS error} = \left[ \frac{1}{n} \sum \varphi^2 \right]^\frac{1}{2} $$

The test data (adapted from GeoNet's) is available on HuggingFace

... And here are the results

2. A demo web app that allows you to run predictions on image and video files while using these models