Current new view synthesis methods have difficulty in recovering high frequency information, resulting in artifacts and blurring in the generated images. Therefore, this paper proposes a new view synthesis method for dense views based on multi-view feature fusion network. Firstly, an anti-aliasing coding algorithm based on FFT is used to capture high-frequency details and improve the quality of the virtual view. Secondly, a multi-scale feature fusion network is introduced and Transformer is used as an aggregation function to extract feature information of different scales by improving the multi-head attention mechanism, and a prior residual module is introduced to alleviate the gradient disappearance problem, so as to enhance local features and improve synthesis quality. Finally, the color monitor function is introduced to implicitly represent the color distribution of the scene to solve the color distortion problem. Experiments show that this method can render high quality virtual view and output high quality 3D image quickly, and the performance is better than other methods.
在提升图像质量方面,基于预训练的MVSNeRF方法[20]通过将预先提取的输入视图图像特征映射到三维体绘制中,从而显著提升了图像质量。NeRF in the Wild(NeRF-W)方法[21]通过共享稀疏视图集合中的NeRF参数,提升了虚拟视点的质量。Mip-NeRF模型[22]采用抗锯齿锥形截锥体(Anti-aliased conical frustums)渲染技术,有效减少了虚拟视点中的锯齿伪影,并显著提高了NeRF在细节表现方面的能力。
ChanS, ShumH Y, NgK T. Image-based rendering and synthesis[J]. IEEE Signal Processing Magazine, 2007, 24(6): 22-33.
[2]
MartínezC M, JavidiB. Fundamentals of 3D imaging and displays: a tutorial on integral imaging, light-field, and plenoptic systems[J]. Advances in Optics and Photonics, 2018, 10(3): 512-566.
[3]
RemondinoF, El-HakimS. Image-based 3D modelling: a review[J]. The Photogrammetric Record, 2006, 21: 269-291.
[4]
MolnarS, CoxM, EllsworthD, et al. A sorting classification of parallel rendering[J]. IEEE Computer Graphics and Applications, 1994, 14(4): 23-32.
[5]
MorelandK D. IceT Users' Guide And Reference[M]. Albuquerque: Sandia National Laboratories, 2011.
[6]
MorelandK, KendallW, PeterkaT, et al. An image compositing solution at scale[C]∥Proceedings of 2011 International Conference for High Performance Computing, Networking, Storage and Analysis, New York, 2011: No.25.
[7]
DietrichA, GobbettiE, YoonS E. Massive-model rendering techniques: a tutorial[J]. IEEE Computer Graphics and Applications, 2007, 27(6): 20-34.
[8]
WaldI, SlusallekP. State of the art in interactive ray tracing[C]∥Proceedings of the 22nd Annual Conference of the European Association for Computer Graphics, Manchester, UK, 2001: 1-21.
[9]
WaldI, MarkW R, GüntherJ, et al. State of the art in ray tracing animated scenes[J]. Computer Graphics Forum, 2009, 28(6): 1691-1722.
[10]
AnantrasirichaiN, GeravandM, BraendlerD, et al. Fast depth estimation for view synthesis[C]∥Proceedings of the 28th European Signal Processing Conference, Amsterdam, The Netherlands, 2021: 575-579.
[11]
HalleM. Multiple viewpoint rendering[C]∥SIGGRAPH'98: Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques, Orlando, USA, 1998: 243-254.
[12]
El-HakimS F. A flexible approach to 3D reconstruction from single images[C]∥ACM SIGGRAPH, Los Angeles, USA, 2001: 12-17.
[13]
ZhangL, DugasP G, SamsonJ S, et al. Single-view modelling of free-form scenes[J]. The Journal of Visualization and Computer Animation, 2002, 13(4): 225-235.
[14]
Cohen-OrD, RichE, LernerU, et al. A real-time photo-realistic visual flythrough[J]. IEEE Transactions on Visualization and Computer Graphics, 1996, 2(3): 225-265.
[15]
ChenS E, WilliamsL. View interpolation for image synthesis[C]∥Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, Anaheim, USA, 1993: 279-288.
[16]
SeitzS M, DyerC R. View morphing[C]∥Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, New Orleans, USA, 1996: 21-30.
[17]
LaveauS, FaugerasO D. 3-D scene representation as a collection of images[C]∥Proceedings of 12th International Conference on Pattern Recognition, Jerusalem, Israel, 1994: 689-691.
[18]
OliveiraQ D, SilveiraT L D, WalterM, et al. A hierarchical superpixel-based approach for DIBR view synthesis[J]. IEEE Transactions on Image Processing, 2021, 30: 6408-6419.
[19]
MildenhallB, SrinivasanP P, TancikM, et al. NeRF: representing scenes as neural radiance fields for view synthesis[C]∥European Conference on Computer Vision, Glasgow, UK, 2020: 405-421.
[20]
ChenA P, XuZ X, ZhaoF Q, et al. MVSNeRF: fast generalizable radiance field reconstruction from multiview stereo[C]∥Proceedings of 2021 IEEE/CVF International Conference on Computer Vision, Montrea, Canada, 2021: 14104-14113.
[21]
MartinB R, RadwanN, SajjadiM S M, et al. NeRF in the wild: neural radiance fields for unconstrained photo collections[C]∥Proceedings of 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 7206-7215.
[22]
BarronJ T, MildenhallB, VerbinD, et al. Mip-NeRF 360: unbounded anti-aliased neural radiance fields[C]∥Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 5460-5469.
[23]
LindellD B, MartelJ N P, WetzsteinG. Autoint: automatic integration for fast neural volume rendering[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2021: 14556-14565.
[24]
LiuL, GuJ, ZawL K, et al. Neural sparse voxel fields[J]. Advances in Neural Information Processing Systems, 2020, 33: 15651-15663.
[25]
GarbinS J, KowalskiM, JohnsonM, et al. Fastnerf: high-fidelity neural rendering at 20 fps[C]∥Proceedings of the IEEE/CVF International Conference on Computer Vision, Nashville, USA, 2021: 14346-14355.
[26]
MüllerT, EvansA, SchiedC, et al. Instant neural graphics primitives with a multiresolution hash encoding[J]. ACM Transactions on Graphics, 2022, 41(4): 1-15.
[27]
KajiyaJ T, VonH B P. Ray tracing volume densities[J]. ACM SIGGRAPH Computer Graphics, 1984, 18(3): 165-174.
[28]
VaswaniA, ShazeerN, ParmarN, et al. Attention is all you need[C]∥Advances in Neural Information Processing Systems, Long Beach, USA, 2017: 6000-6010.