The goal of 3D Object Retrieval research is to ﬁnd effective and effcient similarity functions to allow searching, ranking and clustering of 3D object data for similarity. This is a challenge as 3D Objects typically comprise many different shapes and visual appearance properties, in addition to different material and functional properties. Furthermore, the notion of similarity is often subjective or domain-speciﬁc, hence in many cases can only be approximated. Our work falls into the following areas:
Fig. 1: Process model for feature-based 3D object retrieval [BKSS07]. Often, view-,
surface- or volumetric features are considered.
Fig. 2: 3D retrieval by example object [BS09] (left) or user sketch [YSSK10] (right).
We explore feature-based approaches for global 3D similarity search. In [BKS+05, BKSS07,BS09], we introduce a process model for 3D feature extraction, as well as a range of view-,surface and volume-based features. While view-based features often show to be robust, it is desirable to explore the space of additional features, e.g., based on orientation [SWS10], or by considering global and local features together [SBS13b, BSW+12], or heuristics to form query-dependent weighting of features [BKS+04]. We also explore features useful for 3D retrieval based on user sketches [YSSK10, YYS15]. Sketch-based search is an alternative when e.g., no3D query example object is available.
Fig. 1: Evaluation assesses effectiveness and efficiency of 3D retrieval methods based on benchmark data.
Due to the large design space of 3D search methods, and the complex concept of similarity, experimental evaluation plays an important role in 3D retrieval. We participate regularly in the International Shape Retrieval Contest SHREC. Recently, we evaluated e.g., the scalability of retrieval methods for large repository sizes [SBS+15, LLL+14], including a methodology to create synthetic benchmark data for retrieval evaluation for range image data [SMB+14]. Also, recently we surveyed and compared a substantial number of techniques for sketch-based retrieval [LLG+14]. It is also interesting to compare the performance of different feature-based distance functions together with alternative features [GLS+15]. From the latter, we may be able to derive rules to improve the retrieval effectiveness.
Fig. 1: In the Archeology domain, there exist large repositories of fragmented artifacts (left). Methods based on 3D similarity may help to restore some of these artifacts (right).
Fig. 2: Identiﬁcation of protruding points on an incomplete 3D shape (left). Candidate symmetry planes can be found based on these generating points and applied forshape completion [GSP+14] (middle, right).
Within the EU Project PRESIOUS we are exploring similarity-based approaches for restoration of fractured, eroded or incomplete 3D representations of artifacts from the Cultural Heritage / Archeology domain. In (GCH [GSP+14]), we describe a process model to restore 3D Archaeological object data, starting from reassembly of fragmented parts, over to symmetry- and search-based completion, to eventually provide a plausibly restored object. Symmetry-based restoration in our case must deal with the problem of incomplete input data. We ﬁnd that an approach based on local interest points detected by heat diffusion analysis, can provide good results for detecting planes of symmetry [SGS14]. These can then be exploited to complete formerly incomplete objects. Recent results of application of the whole pipeline can be found in [AGS+15].
Evaluation of the performance of similarity-based restoration is a challenge, as often, there is no ground truth available which speciﬁes the original shape before fragmentation and erosion took place. To this end, we proposed a ﬁrst automatic 3D fragmentation process, which can create synthetic fracture benchmark data [GBS+15].
Figure 1: User query interfaces for 3D building data [BBK+09] based on 3D sketch editing (left) and room connectivity (right).
Another important application domain for 3D objects is in Architecture. Nowadays, many architects design buildings and other structures using 3D technology. This gives rise to the question, how 3D architecture models can be stored and retrieved in 3D architecture repositories. With the project PROBADO, together with two large national German research libraries, we have researched archival and retrieval of building models. In [BBC+10], we have summarized technical and organizational problems and solutions when setting up a 3D Digital Library for architectural data. The corresponding library service can follow a distributed web-based architecture, and in [BKHS09] we have discussed design choices made in PROBADO. Regarding user access, in [BBK+09] we have proposed a variety of domain-adapted user query modalities. These include 3D sketch editing for building shapes, as well as an abstract query interface based on room connectivity structures or ﬂoor plans.
Figure 1: Visualization of rankings [SBW09] (left) and clusterings [BKPS04] (right) provide user access to large repositories.
Besides iterative query speciﬁcation, also the visual exploration of 3D object repositories is important for accessing data. Cluster visualization methods like the Self-Organizing Map algorithm [BKPS04] or rank-based techniques [SBW09], together with suitable 3D feature vector representations can be useful for overviewing and navigating large 3D repositories.