Efficient Random Access Compression Schemes for Image-Based Representations
Public Deposited- Creator
-
Pratapa, Srihari V
- Affiliation: College of Arts and Sciences, Department of Computer Science
- Abstract
- In the last two decades, a new class of rendering methods called image-based rendering has emerged as an alternative to traditional geometry-based rendering methods. Image-based representations require an intensive amount of image data, making compression a critical issue. In particular, we need efficient compression schemes that map well to the current graphics hardware and provide random access capability for interactive applications. In this dissertation, we present new, efficient random-access compression schemes for three different image-based representations. Our first compression scheme, MPTC, introduces the idea of transcoding compressed video data directly to compressed textures such that the decoded output can be rendered using commodity texture mapping hardware. Our method improves rendering speeds up to an order of magnitude on mobile devices. Next, we propose two novel compression schemes, RLFC and HMLFC, for encoding light field images. In the RLFC algorithm, we describe a new hierarchical approach that provides progressive streaming and decoding capabilities that enable interactive rendering of light fields. HMLFC combines two different approaches for encoding light fields, motion compensation and hierarchical approaches to capture redundancies in a hybrid fashion; the compression rates with the hybrid approach improve by a factor of 2-5 times over the underlying motion and hierarchical schemes. Our third compression method, RANDM, deals with compressing depth maps using a range-partitioning and a global dictionary. RANDM is the first compression scheme that enables random access decoding for depth maps and achieves compression rates similar to the existing schemes. Each of these three methods allow compressed representation of image data in memory while rendering and selectively decoding required data quickly during run-time. We also present a novel, deep learning-based approach (neural networks), TexNN, for predicting the right set of encoding parameters for fast encoding of textures into compressed texture formats. Our method offers up to a magnitude of speed-up compared to prior compression algorithms while maintaining similar compression quality.
- Date of publication
- 2020
- Keyword
- DOI
- Resource type
- Advisor
- Manocha, Dinesh
- Olson, Tom
- Lin, Ming C
- Bishop, Gray
- Mayer-Patel, Ketan
- Degree
- Doctor of Philosophy
- Degree granting institution
- University of North Carolina at Chapel Hill Graduate School
- Graduation year
- 2020
- Language
Relations
- Parents:
This work has no parents.
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Pratapa_unc_0153D_19144.pdf | 2020-06-22 | Public | Download |