3D Intrinsic scene characteristics extraction framework for a single image
Three-Dimensional (3D) shape reconstruction is an important area of computer vision research because it has numerous potential applications from entertainment production to industrial inspection and clinical analysis. Existing 3D Intrinsic Scene Characteristics (3D-ISCs) extraction methods for a sin...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/id/eprint/18570/1/3D%20Intrinsic%20Scene%20Characteristics%20Extraction%20Framework%20For%20A%20Single%20Image%2024%20Pages.pdf http://eprints.utem.edu.my/id/eprint/18570/2/3D%20Intrinsic%20scene%20characteristics%20extraction%20framework%20for%20a%20single%20image.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Three-Dimensional (3D) shape reconstruction is an important area of computer vision research because it has numerous potential applications from entertainment production to industrial inspection and clinical analysis. Existing 3D Intrinsic Scene Characteristics (3D-ISCs) extraction methods for a single image have focused solely on estimating diffuse characteristics, i.e. 3D shape, illumination, and reflectance models, of an object. As a result, they have neglected the specular characteristic, the shiny areas of a glossy surface. In reality, many real-world objects emit both specular and diffuse reflections, and thus the specular component may decrease the performance of the 3D-ISCs methods. This study has developed a framework to extract all of these characteristics. The framework combines a Specular Removal (SR) method and a Shape, Illumination, and Reflectance From Shading (SIRFS) method under a Bidirectional Reflectance Distribution Function (BRDF) model. Since the previous SR methods suffered from hue-saturation ambiguity, they are not suitable for this framework. To solve this problem, two SR methods were developed, evaluated, and compared with the standard SR methods. The proposed SR methods are referred as Chaotic Segmentation (CS) and Sparse Coding (SC) methods. To combine the SR and SIRFS methods, two BRDF models were also developed, evaluated, and compared. These models are referred as Modified Dichromatic Reflectance (MDR) and Modified Blinn-Phong (MBP) models. The performances of the proposed SR methods and the BRDF models for extracting 3D-ISCs were evaluated based on public datasets. The results showed that the SC method was more satisfactory compared to the CS and the benchmark method (iterative method). The accuracies of the diffuse and specular characteristics were improved by 7.6% and 53.5% respectively. Moreover, the combination of SC method and MDR model was capable of outperforming the SIRFS method. The computational speed was 19.2% faster. Meanwhile, the average accuracies of depth, surface normal, illumination, shading, and reflectance were improved by 11.4%, 6.5%, 50.5%, 35.2%, and 5.1% respectively. This study indicates that the specular reflection is an important aspect of
3D reconstruction from a single image. The proposed framework has also made considerable improvements in terms of accuracy and computational time of extracting 3D-ISCs. |
|---|