Mr. Georgios Albanis and Mr. Vasileios Gkitsas presented their works to the OmviCV Workshop of CVPR 2021:
- > G. Albanis, N. Zioulis, P. Drakoulis, V. Gkitsas, V. Sterzentsenko, F. Alvarez, D. Zarpalas, P. Daras, “Pano3D: A Holistic Benchmark and a Solid Baseline for 360 Depth Estimation”, in IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June, 2021.
- > V. Gkitsas, V. Sterzentsenko, N. Zioulis, G. Albanis, D. Zarpalas, “PanoDR: Spherical Panorama Diminished Reality for Indoor Scenes”, in IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June, 2021.
The recorded version of the workshop can be found here. The Pano3D presentation starts at 2:30:29, while the PanoDR presentation starts at 7:17:40.
Mr. Albanis’ article is about Pano3D, a new benchmark for depth estimation from spherical panoramas which aims to assess performance across all depth estimation traits, the primary direct depth estimation performance targeting precision and accuracy, and also the secondary traits, boundary preservation and smoothness. Pano3D moved beyond typical intra-dataset evaluation to inter-dataset performance assessment. By disentangling the capacity to generalize in unseen data into different test splits, Pano3D represented a holistic benchmark for 360 degree depth estimation. Pano3D was used as a basis for an extended analysis seeking to offer insights into classical choices for depth estimation and resulted into a solid baseline for panoramic depth that follow-up works can built upon to steer future progress.
Mr. Gkitsas’ work focuses on structure-aware counterfactual inpainting. A model is proposed that initially predicts the structure of an in-door scene and then uses it to guide the reconstruction of an empty – background only – representation of the same scene. Diminished Reality (DR) fulfills the requirement of applications related to interior space re-design and helps removing existing objects in the scene, essentially translating this to a counterfactual inpainting task. The trained model compared against other state-of-the-art methods showed superior results in both quantitative metrics and qualitative results, but more interestingly, the proposed approach exhibited a much faster convergence rate.
Both works were supported by the EC funded H2020 project ATLANTIS.