CORSAIR: Convolutional Object Retrieval and Symmetry-AIded Registration

Abstract

This paper considers online object-level mapping using partial point-cloud observations obtained online in an unknown environment. We develop and approach for fully Convolutional Object Retrieval and Symmetry-AIded Registration (CORSAIR). Our model extends the Fully Convolutional Geometric Features model to learn a global object-shape embedding in addition to local point-wise features from the point-cloud observations. The global feature is used to retrieve a similar object from a category database, and the local features are used for robust pose registration between the observed and the retrieved object. Our formulation also leverages symmetries, present in the object shapes, to obtain promising local-feature pairs from different symmetry classes for matching. We present results from synthetic and real-world datasets with different object categories to verify the robustness of our method.

Reconstruction of Architectural Heritage With Symmetrical Components

“…Therefore, finding symmetries in such geometric data is a crucial topic in geometry processing. Extracted symmetry information has been used for many applications, such as shape matching (Sharma and Ovsjanikov, 2021), retrieval (Zhao et al, 2021), geometry completeness (Schiebener et al, 2016), structural and façade accuracy (Harshit et al, 2021, Zhou et al, 2016, and procedural modelling (Rumezhak et al, 2021).…”

Accurate Instance-Level CAD Model Retrieval in a Large-Scale Database

“…Dahnert et al [10] segment the foreground from the input scan and complete the segmented objects by stacking hourglass encoder-decoders, resulting in a joint embedding of 3D scans and CAD models. More recently, CORSAIR [9] extends FCGF [23] to simultaneously learn a global feature for retrieval as well as local point features for alignment.…”

“…An interesting variant of CD has been introduced in [9], called the Single-direction Chamfer Distance (SCD). It simply drops one side of squared distances in (1), resulting in…”

“…Current methods, such as [1][8], concentrate on computing the 9DoF alignment between scanned objects and CAD models while only requiring the retrieval of a same-category model due to the usage of a small ground truth pool. Zhao et al [9] and Dahnert et al [10] claim retrieval of CAD instances with higher similarity, but they still constrain the model search to a relatively small database containing an augmented ground truth set for each scanned object. We instead focus on instance-level retrieval of CAD models in a large shape database, which is more practical in real-world scenarios.…”

“…It relies on a feature-based k-nearest neighbor search followed by a geometrybased re-ranking, which can considerably improve the quality of retrieved CAD models. • We introduce a variant of the Single-direction Chamfer Distance [9] that is more robust to outliers when measuring point set distances between partial scans and CAD models. As demonstrated by our experiments, the combination of geometric re-ranking outperforms the plain application of state-of-the-art feature embedding networks.…”