Do We Need a New Large-Scale Quality Assessment Database for Generative
Inpainting Based 3D View Synthesis? (Student Abstract)
Sadbhawna, Vinit Jakhetiya, Badri N. Subudhi, Harshit Shakya, Deebha Mumtaz
Indian Institute of Technology Jammu, India
{2018rcs0013, vinit.jakhetiya, subudhi.badri, 2020pcs2064, 2018rcs0011}@iitjammu.ac.in
Abstract
The advancement in Image-to-Image translation techniques
using generative Deep Learning-based approaches has shown
promising results for the challenging task of inpainting-based
3D view synthesis. At the same time, even the current 3D
view synthesis methods often create distorted structures or
blurry textures inconsistent with surrounding areas. We ana-
lyzed the recently proposed algorithms for inpainting-based
3D view synthesis and observed that these algorithms no
longer produce stretching and black holes. However, the ex-
isting databases such as IETR, IVC, and IVY have 3D-
generated views with these artifacts. This observation sug-
gests that the existing 3D view synthesis quality assessment
algorithms can not judge the quality of most recent 3D syn-
thesized views. With this view, through this abstract, we an-
alyze the need for a new large-scale database and a new
perceptual quality metric oriented for 3D views using a test
dataset.
Introduction
A good 3D-synthesized images/videos can provide con-
sumers with a more engaging and better immersive ex-
perience. Free Viewpoint Video (FVV), 3D-Television,
360°video, Virtual Reality (VR) are some of the applications
of 3D-synthesis, famous because of their realistic and inter-
active experience (Shih et al. 2020; Niklaus et al. 2019). Un-
fortunately, the rendered 3D views, even using the contem-
porary methods, cannot generate the perfect novel 3D view
(Shih et al. 2020). These methods cannot perform efficiently
on complex surfaces and produce some artifacts, as shown in
Fig. 1. The artifacts in the 3D synthesized views are differ-
ent from the conventional artifacts in regular natural images.
With the advancement of efficient algorithms for generating
3D synthesized views, it is required to have an image quality
assessment (IQA) algorithm which can automatically judge
the perceptual quality of generated 3D synthesized views
that match with the human visual system. The 3D IQA al-
gorithms can judge the perceptual quality and are also help-
ful in the fast development of Image Restoration (IR) (IR
includes tasks such as super-resolution (SR), denoising, en-
hancement, etc.) algorithms. With this view, through this ab-
stract, we will address the research problem statement that is
Copyright © 2022, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Figure 1: (a). A synthesized view. (b). The failure (green
arrows) of a random patch (red window) in a 3D synthesized
view. Synthesized Using: (Shih et al. 2020)
there a need for creating a large-scale 3D synthesized IQA
database which is generated using the recently proposed 3D
synthesized view generation algorithms?
Related Work
Novel View Synthesis(NVS): From proxy geometry based
traditional methods to Generative Adversarial Networks
(GANs) based deep learning methods, NVS has evolved
with the advancement in computer vision techniques. (Shih
et al. 2020; Niklaus et al. 2019) are some of the new view
synthesis techniques. Synthesizing novel views requires
stipulations such as comprehensive scene understandings,
preserving structures observed in the input semantics,
lowered baseline requirements, etc. Even the newest view
synthesis methods in the literature lack in one or the other
of these stipulations. Hence, much research is being done to
make the view synthesis perfect. Figure 1 shows an example
of the failure of such a contemporary 3D synthesized
method. From this Figure, it is clear that view synthesis
does not render a clear view in certain circumstances, such
as complex surfaces.
Image Quality Assessment(IQA): SSIM(Wang et al.
2004) is the most widely used IQA method as it introduced
the structural similarity in comparing images as compared
to the Mean Square Error (MSE) value, which are FR (Full-
Reference) IQAs. There are various other popular NR (No-
Reference) IQAs, such as BRISQUE, NIQE, etc., in the lit-
erature. There are various new IQAs based on deep features
The Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22)
13039