3D Reconstruction of real-world scenes is a field that has been gaining momentum in recent
years and that is due to its many uses in industry in a lot of domains: medicine, sports,
autonomous vehicles, architecture, interior design, construction, etc. In our project, we read
and worked with recent articles from the last few years to try and investigate the most
advanced methods to create a 3D Reconstruction using Machine Learning, then, after
comparison between the methods, we had to pick one of them, learn more thoroughly about
it then think about ways to improve its results. We’ve picked Mildenhall, B., Srinivasan, P.
P., Tancik, M., Barron, J. T., Ramamoorthi, R., & Ng, R. (2021). Nerf: Representing
scenes as neural radiance fields for view synthesis. Communications of the ACM,
65(1), 99-106. This article was written by researchers from UC Berkeley, Google research
and UC San Diego. Plus, it was quoted 2056 times.
In order to enhance the results, we wanted our model to be robust to different variations of
the input images so we gave the model a randomly augmented training set and checked the
results. For that to be done, we used TensorFlow to perform data augmentation on the
training set.
We tried two data augmentation techniques on the training set:
1. Adding Gaussian noise with different energy (standard variation)
2. Flipping the image right to left
We’ve concluded that the model is robust to real-world “noisy” images since we’ve got great
results and we even can say that after adding noise, we’ve got a 3D that is a little more “nice
to look at”. On the other hand, flipping the image right to left (180 degrees) has worsened the
results.
