Nikhil N.A.
Praveen C.
Jagadish B.
Vijayalaxmi P.
Download our project report
Take a look at our code
Model Overview
We propose a design, wherein we use the classical image warping based on camera parameters, we mainly extract the camera’s intrinsic parameters after the process of calibrating the camera. With the extracted intrinsic matrix of the given camera, we then generate a disparity map for the given scene whose novel view is to be generated. Once the disparity map is generated, we then render it to a point cloud, transform it, and then project the transformed point cloud to a 2D image. This projected image will of course contain holes, which we then fill after generating a custom mask. To fill the holes, we use inpainting techniques. Navier-Stokes, and Telea methods to be specific.
Results
The results shown in above figure represents the 2D projected image, the respective mask, the results obtained off Navier-Stokes algortihm and the later image shows results obtained from inpainting off Telea algorithm. Both the results look about the same with some minor differences, indicating the generation of novel view, for the given scene.
With the proposed method, we were able to achieve:
• Peak Signal To Noise Ratio (PSNR) of about 28.8dB
• Structural Similarity Index Measure (SSIM) of about 0.58
Abstract
The rapid advancements in the field of Augmented Reality and Virtual Reality and the vastly available resources make it conceivable to join hands in innovation. Smart gadgets nowadays such as mobiles, tablets, AR devices etc, as we even speak of it, are highly capable of rendering various things in augmented reality and virtual reality as well. This has led to the development and advancements in 3D reconstruction and Novel View Generations of any given object. With this motivation, we propose a design, to reconstruct or generate a new view, for a given object. In this project we mainly aim to generate a new view, for any given image/object, employing computer vision methods and algorithms. To generate a novel view, we take an input sparse view, calibrate the camera, and then render the subject to a 3-dimensional Point Cloud, and select a suitable view to project to an image. We then employ masking technique, and fill the holes, generated using Navier Stokes and Telea algorithms. We then present the results, of using the described method by generating a new view of the given input image of a man.