완전한 3차원 물체 복원을 위한 실시간 다음 최적 시점 선택 = Real-time Camera Next-Best-Viewpoint Selection for Complete 3D Object Reconstruction

국문 초록 (Abstract)

본 연구는 3D 물체 복원 과정에서 발생하는 기존 Next-Best-View (NBV) 시스템의 한계점을 해결하고, 보다 효율적이고 정확한 3D 복원을 위한 새로운 방법론을 제시한 다. 기존 NBV 시스템들은 ray 샘플링이나 개수 제한을 통한 불확실성 계산, 고정된 view-space에서의평가등으로인해 free-space에서의불확실성을일반화하기어렵다는 문제점을 가지고 있었다. 이러한 한계를 극복하기 위해 본 연구에서는 Distance Guidance Regularization와 Unknown Voxels Count Regularization의 두 가지 핵심 보정 메커니즘을 제안한다. Distance Guidance Regularization는 물체와의 거리에 따른 가중치를 도입하여 지나 치게 가까운 시점 선택을 제한하며, Unknown Voxels Count Regularization는 미관측 복셀의 수를 고려하여 보다 적응적인 시점 선택을 가능하게 한다. 또한, 기존 NBV 시스템의 계산 복잡성 문제를 해결하기 위해 Orientation-First Search라는 새로운 후보 시점 탐색 방법을 도입하였다. 이 방법은 방향과 거리의 탐색을 분리하여 순차적으로 수행함으로써, 계산 효율성을 크게 향상시켰다. 특히, 단일 반구 상의 후보 방향에 대해 먼저정보 이득을 계산한 후, 선택된 방향에 대해서만 다양한 거리를 평가하는 방식을 통해 효율적인 시점 선택이 가능하도록 하였다. 실험 결과, 제안된 시스템은 넓은 후보 공간에서 시점을 선택했을 때 기존 방법과 비교하여 전체 복원 완성도가 약 6% 높은 수치를 보인다. 이는 형태 완성도가 약 10% 높은 수치를 보여 적절한 거리 유도를 통해 물체의 완성도를 향상시켰다. 그리고 Orientation-First Search를 통해 후보 시점 탐색 시간을 약 9배 향상시키며 방향각에 대한 균일한 선택을 하게 하여 복원 완성도도 향상시킬 수 있음을 확인하였다.

목차 (Table of Contents)

• 요약································································································································· i
• 표목차······························································································································ ii
• 그림목차·························································································································· ii
• 1. Introduction················································································································ 1
• 1.1. Motivation············································································································ 1
• 1.2. Objectives············································································································ 2
• 2. Related Works········································································································ 3
• 2.1. Next-Best-View(NBV) Selection······································································ 3
• 2.2. 3D Object Reconstruction Pipeline································································ 4
• 2.3. Volumetric Information Gains········································································· 6
• 3. Limitation of NBV Selection and Criteria··························································· 8
• 3.1. Limitation of Candidate View Space····························································· 8
• 3.2. Limitation of Volumetric Information Gains················································ 8
• 4. Real-time Next-Best-Viewpoint Selection························································· 13
• 4.1. Problem Formulation························································································· 13
• 4.2. Next-Best-Viewpoint Criteria with Regularization··································· 13
• 4.2.1. Distance Guidance Regularization························································· 13
• 4.2.2. Unknown Voxels Count Regularization················································ 14
• 4.3. Next-Best-Viewpoint Selection····································································· 15
• 4.3.1. Brute-force Search··················································································· 15
• 4.3.2. Orientation-First-Search·········································································· 15
• 4.4. Real-time Visualization··················································································· 17
• 5. Experiments············································································································· 19
• 5.1. Datasets·············································································································· 19
• 5.2. Evaluation Metrics··························································································· 19
• 5.3. Experiments······································································································ 20
• 5.3.1. Experiment for Regularizations····························································· 21
• 5.3.2. Experiment for Weight Parameter······················································· 21
• 5.3.3. Experiment for Search Methods···························································· 24
• 5.4. Visualization······································································································ 25
• 6. Conclusion················································································································ 26
• 참고문헌························································································································ 27
• 영문초록(Abstract)······································································································· 29
• 감사의글······················································································································ 31