Title Page
Contents
ABSTRACT 8
Ⅰ. Introduction 10
Ⅱ. Background 15
2.1. Neural Radiance Fields (NeRF) 15
2.2. Depth Supervision NeRF 16
2.3. NeRF Optimizing Camera Poses 17
2.4. Bundle Adjusting Neural Radiance Fields 19
Ⅲ. Methodology 20
3.1. Pose Estimation using Motion Priors 20
3.2. Optimizing with Depth 21
3.3. Deriving Depth-Supervision Sampling 24
Ⅳ. Experiments 26
4.1. Evaluation Metrics 26
4.2. Baseline Comparison 27
4.3. Ablation Study 34
Ⅴ. Conclusion 40
REFERENCES 41
ABSTRACT IN KOREAN 46
Table 1. Quantitative results on 5 input views 33
Table 2. Rotation matrix difference (RMD) in training datasets 38
Table 3. Quantitative analysis on camera pose optimization 39
Figure 1. Framework of CM-NeRF. 14
Figure 2. Overview of our CM-NeRF pipeline. 25
Figure 3. Qualitative comparisons in real datasets using five input views. 31
Figure 4. Visualize optimized camera pose comparison. 36
Figure 5. Qualitative comparison of pose and depth ablation 37