Projects Last Projects

We present an approach to training stereo neural networks
(NNs) using novel views synthesized from 3D Gaussian Splatting (GS)
models. Traditional stereo datasets require precise camera calibration
and advanced equipment, making them expensive to create. Our method
overcomes this limitation by generating novel views using Gaussian splatting,
a 3D scene reconstruction and rendering technique. By augmenting
training data with synthetic stereo pairs at varying perspectives, we reduce
the cost of creating the dataset and improve the model’s ability to
fit the learned task. We evaluated our approach on the DTU [1] dataset
(standard stereo benchmark) and demonstrated improvements in depth
estimation accuracy and 3D nuances compared to models trained on
conventional datasets. Our findings suggest that integrating Gaussian
Splatting-based novel view synthesis into stereo training pipelines can
reduce training costs and enhance performance across real-world and
synthetic stereo tasks.

Please, see project report.

Please, see final presentation.