Chapter 13: 3D reconstruction

• The DiLiGenT photometric stereo dataset provides images taken under calibrated directional lighting and objects with general reflectance along with ground truth shapes (Shi, Mo et al. 2019). It also provides a taxonomy and evaluation of photometric stereo methods for general non-Lambertian materials and unknown lighting.

• NYU3D (Silberman, Hoiem et al. 2012) and ScanNet (Dai, Chang et al. 2017) were some of the early 3D indoor scene datasets used to study 3D reconstruction and range fusion algorithms. More recent algorithms such as Chabra, Lenssen et al. (2020) or Weder, Schonberger et al. (2021) use some combination of 3D Scenes (Zhou and Koltun 2013), ICL-NUIM (Handa, Whelan et al. 2014), ShapeNet (Chang, Funkhouser et al. 2015), and Tanks and Temples (Knapitsch, Park et al. 2017). Reviews of RGB-D datasets can be found in Firman (2016) and Zollhöfer, Stotko et al. (2018).

C.2 Software

• Over the years, a number of 3D human body and motion datasets have been captured, including HumanEva (Sigal, Balan, and Black 2010), MPI FAUST (Bogo, Romero et al. 2014), Panoptic Studio (Joo, Simon et al. 2019), EHF (Pavlakos, Choutas et al. 2019), AMASS (Mahmood, Ghorbani et al. 2019), and 3D Poses in the Wild (3DPW) (von Marcard, Henschel et al. 2018). $ ^{2} $

• In parallel with these datasets, 3D human body models and fitting algorithms have been developed, including SCAPE (Anguelov, Srinivasan et al. 2005), BlendSCAPE (Hirshberg, Loper et al. 2012). SMPL (Loper, Mahmood et al. 2015), MANO (Joo, Simon, and Sheikh 2018), SMPL-X (Pavlakos, Choutas et al. 2019), VIBE (Kocabas, Athanasiou, and Black 2020), ExPose (Choutas, Pavlakos et al. 2020), STAR (Osman, Bolkart, and Black 2020), Learned Gradient Descent (Song, Chen, and Hilliges 2020), and FrankMoCap (Rong, Shiratori, and Joo 2020). These are described in more detail in Section 13.6.4.