Date of Award
Electrical and Computer Engineering
3D, Multi-frame, Multi-view, Reconstruction, Stereo, Super resolution
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Multi-view stereo is a popular method for 3D-reconstruction. Super resolution is a technique used to produce high resolution output from low resolution input. Since the quality of 3D-reconstruction is directly dependent on the input, a simple path is to improve the resolution of the input.
In this dissertation, we explore the idea of using super resolution to improve 3D-reconstruction at the input stage of the multi-view stereo framework. In particular, we show that multi-view stereo when combined with multi-frame super resolution produces a more accurate 3D-reconstruction.
The proposed method utilizes images with sub-pixel camera movements to produce high resolution output. This enhanced output is fed through the multi-view stereo pipeline to produce an improved 3D-model. As a performance test, the improved 3D-model is compared to similarly generated 3D-reconstructions using bicubic and single image super resolution at the input stage of the multi-view stereo framework. This is done by comparing the point clouds of the generated models to a reference model using the metrics: average, median, and max distance. The model that has the metrics that are closest to the reference model is considered to be the better model.
The overall experimental results show that the generated models, using our technique, have point clouds with average mean, median, and max distances of 4.3\%, 8.8\%, and 6\% closer to the reference model, respectively. This indicates an improvement in 3D-reconstruction using our technique. In addition, our technique has a significant speed advantage over the single image super resolution analogs being at least 6.8x faster.
The use of multi-frame super resolution in conjunction with the multi-view stereo framework is a practical solution for enhancing the quality of 3D-reconstruction and shows promising results over single image up-sampling techniques.
Lee, Michael, "Enhancing Multi-View 3D-Reconstruction Using Multi-Frame Super Resolution" (2022). Electronic Theses and Dissertations. 9025.