Indoor Image Surface Normal Estimation Using Convolutional Neural Network

다국어 초록 (Multilingual Abstract)

A surface normal vector is related to material of object, shape of geometry, re-
flection of light. To estimate other features, the surface normal vector is necessary.
Therefore, the surface normal estimation problem is one of the most important topic
in inverse graphics.
Previously, computing surface normal image from single RGB-D image has been stud-
ied since estimating surface normal vector from single RGB image is extremely ill-
posed problem. Recently, deep learning based approach has been presented for the
task. However, it is not well-known that which deep learning model is efficiently solve
the problem. Also there are limitations in dataset computed from real RGB-D images.
In this thesis, first we describes the problems in previous works; the dataset. Sec-
ond, we suggest a way to overcome the limitation of previous work; synthetic dataset.
Lastly, we test several neural net model with set of experiments to find the most proper
model. After that we train the neural network model with synthetic dataset.
The model’s prediction of surface normal image looks plausible. Numerically the
trained model achieve lower error compared to previous works.

목차 (Table of Contents)

• 1 INTRODUCTION 1
• 2 BACKGROUND AND RELATED WORKS 3
• 2.1 Convolutional Network 3
• 2.2 Convolutional Network Nodels 4
• 2.3 Single Image Surface Normal Estimation 7
• 2.4 Synthetic Dataset 8
• 3 METHODS 10
• 3.1 Training Data 10
• 3.2 Preprocessing 13
• 3.3 Training Setup 15
• 3.4 Loss Function 16
• 3.5 Neural Network 17
• 4 EXPERIMENT RESULTS 19
• 4.1 Loss Comparison 21
• 4.2 Model Comparison with Partial Dataset 21
• 4.3 Model Comparison with Full Dataset 24
• 5 CONCLUSION 27
• Abstract (In Korean) 31