SyB3R: A Realistic Synthetic Benchmark for 3D Reconstruction from Images
Benchmark datasets are the foundation of experimental evaluation in almost all vision problems. In the context of 3D reconstruction these datasets are rather difficult to produce. The field is mainly divided into datasets created from real photos with dif
- PDF / 3,605,145 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 28 Downloads / 301 Views
Abstract. Benchmark datasets are the foundation of experimental evaluation in almost all vision problems. In the context of 3D reconstruction these datasets are rather difficult to produce. The field is mainly divided into datasets created from real photos with difficult experimental setups and simple synthetic datasets which are easy to produce, but lack many of the real world characteristics. In this work, we seek to find a middle ground by introducing a framework for the synthetic creation of realistic datasets and their ground truths. We show the benefits of such a purely synthetic approach over real world datasets and discuss its limitations.
1
Introduction and Related Work
The reconstruction of digital 3D models from images includes various tasks ranging from camera calibration, over the determination of camera positions (structure from motion) and dense reconstruction, to surface generation and interpretation (e.g. segmentation). Over the last years, a still rising number of algorithms have been proposed that are able to obtain high-quality 3D reconstructions in several application scenarios, including those where other approaches are not easily applicable. The state of the art of this field is still improving rapidly. An overview about recent advances in structure from motion methods can be found in [1], while [2–4] offer reviews of multi-view stereo algorithms. The need to objectively compare such algorithms and to investigate their intrinsic properties has led to the proposal of many benchmark datasets, which provide reference data (i.e. measured by other sensors) or ground truth (based on synthetic models). Both types of datasets have complementary benefits and limitations. Datasets that are based on real measurements have the advantage that all the effects that can occur during data acquisition are (at least potentially) included as they actually happen during the acquisition. This property of real datasets is of course only theoretical, since the concrete, practical experimental setup is limiting the effects that can be covered. These datasets mostly contain a few example and often simplified scenarios, where images are obtained under Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46478-7 15) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016 B. Leibe et al. (Eds.): ECCV 2016, Part VII, LNCS 9911, pp. 236–251, 2016. DOI: 10.1007/978-3-319-46478-7 15
SyB3R: A Realistic Synthetic Benchmark for 3D Reconstruction
237
Fig. 1. We present a framework for the synthetic generation of realistic 3D reconstruction datasets with ground truth data which allows the evaluation of 3D reconstruction methods in fully controlled and possibly non-standard application scenarios.
fixed conditions (e.g. same lighting, same camera, a certain baseline, etc.). Furthermore, these datasets cannot provide ground truth but only reference data, which is acquired by a sensor (mostly structured light or laser scanning) that is assu
Data Loading...
