NeRF의 정확한 3차원 복원을 위한 거리-엔트로피 기반 영상 시점 선택 기술
Distance and Entropy Based Image Viewpoint Selection for Accurate 3D Reconstruction with NeRF
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper proposes a new approach with a distance-based regularization to the entropy applied to the NBV (Next-Best-View) selection with NeRF (Neural Radiance Fields). 3D reconstruction requires images from various viewpoints, and selecting where to capture these images is a highly complex problem. In a recent work, image acquisition was derived using NeRF's ray-based uncertainty. While this work was effective for evaluating candidate viewpoints at fixed distances from a camera to an object, it is limited when dealing with a range of candidate viewpoints at various distances, because it tends to favor selecting viewpoints at closer distances. Acquiring images from nearby viewpoints is beneficial for capturing surface details. However, with the limited number of images, its image selection is less overlapped and less frequently observed, so its reconstructed result is sensitive to noise and contains undesired artifacts. We propose a method that incorporates distance-based regularization into entropy, allowing us to acquire images at distances conducive to capturing both surface details without undesired noise and artifacts. Our experiments with synthetic images demonstrated that NeRF models with the proposed distance and entropy-based criteria achieved around 50 percent fewer reconstruction errors than the recent work.
Abstract
This paper proposes a new approach with a distance-based regularization to the entropy applied to the NBV (Next-Best-View) selection with NeRF (Neural Radiance Fields). 3D reconstruction requires images from various viewpoints, and selecting where to capture these images is a highly complex problem. In a recent work, image acquisition was derived using NeRF's ray-based uncertainty. While this work was effective for evaluating candidate viewpoints at fixed distances from a camera to an object, it is limited when dealing with a range of candidate viewpoints at various distances, because it tends to favor selecting viewpoints at closer distances. Acquiring images from nearby viewpoints is beneficial for capturing surface details. However, with the limited number of images, its image selection is less overlapped and less frequently observed, so its reconstructed result is sensitive to noise and contains undesired artifacts. We propose a method that incorporates distance-based regularization into entropy, allowing us to acquire images at distances conducive to capturing both surface details without undesired noise and artifacts. Our experiments with synthetic images demonstrated that NeRF models with the proposed distance and entropy-based criteria achieved around 50 percent fewer reconstruction errors than the recent work.
참고문헌 (12)
- S. Isler, "An Information Gain Formulation for Active Volumetric 3D Reconstruction", 2016 IEEE International Conference on Robotics and Automation(ICRA), pp.3477~3484 (2016)
- L. hou, "Volumetric Next Best View by 3D Occupancy Mapping Using Markov Chain Gibbs Sampler for Precise Manufacturing", IEEE Access, 7, pp.121949~121960 (2019)[DOI]
- B. Mildenhall, "NeRF : Representing Scenes as Neural Radiance Fields for View Synthesis", European Conference on Computer Vision, 123456, pp.405~421 (2020)
- S. Lee, "Uncertainty Guided Policy for Active Robotic 3D Reconstruction Using Neural Radiance Fields", IEEE Robotics and Automation Letters, 7(4), pp.12070~12077 (2022)[DOI]
- Y. Ran, "NeurAR : Neural Uncertainty for Autonomous 3D Reconstruction With Implicit Neural Representations", IEEE Robotics and Automation Letters, 8(2), pp.1125~1132 (2023)[DOI]
- L. M. Wong, "Next-best-view algorithm for object reconstruction", Sensor Fusion and Decentralized Control in Robotic Systems, 3523, pp.191~200 (1998)[DOI]
- C. E. Shannon, "A mathematical theory of communication", The Bell System Technical Journal, 27(3), pp.379~423 (1948)[DOI]
- N. Wang, "Pixel2mesh : Generating 3d mesh models from single rgb images", European conference on computer vision, 11215, pp.52~67 (2018)
- T. Muller, "Instant Neural Graphics Primitives with a Multiresolution Hash Encoding", ACM Transactions on Graphics, 41(4), pp.1~15 (2022)[DOI]