3 D OBJECT RETRIEVAL METHODS 3 D Models

3 D OBJECT RETRIEVAL METHODS 3 D Models Multiple View The 3 D object is represented by a virtual 3 D model and it can be created using statistics-, volume-, or surface-geometricbased method. However, it’s computationally expensive, poor performance of reconstruction methods leads in low-quality 3 D models. Representative view selection View capture Have 4 stages The fundamental elements for viewbased 3 D object analysis and consists of a group of a cameras capturing views from different directions This method use a single or multiple and do not required a 3 D model. It is easy to obtain images of real objects and it have better object retrieval performance Feature extraction Object matching using multiple views Based on one-to-one image matching and view based 3 D object retrieval. It focuses on multiple view matching Estimate the relevance among different 3 D objects It provide rich information. However, it introduce redundant and noisy data and result in high computation costs. It difficult to extract features for multiple views and still requires further investigation View selection view capture and representative 1. 2. 3. 4. A compact group of views that can provide adequate and concise information for 3 D object description. The existing method can be divided into 4 paradigms View selection Future extraction Object Matching with multiple views Future Directions Representative view selection from a pool The number of representative views is determined by a predesigned camera array A typical method of view selection from a pool is Adaptive Views Clustering (AVC). It contains 320 views. Predesigned camera array Eg: Petro Daras and Apostolos Axennopolos proposed a compact Multiview descriptor (CMVD) Synthesized view generation • Representation for accurate model attributing (Panorama) method to represent surface information • The spatial structure circular descriptor (SSCD) synthesized view method is invariant to rotation and scaling. Eg: Elevation descriptor (ED) Extracts six range views for each 3 D model from a bounding box. Incremental view selection Selects the representative views from a large pool and involves query view suggestion (QVS) which is using relevance feedback from users. Future Extraction • For multiple view representation Object Matching with multiple views • Methods to integrate distance from a view pairs across two objects. • Limited by computational costs, and for a large-scale 3 D object database, it is hard to learn the global information using only one big graph structure. • Includes extracting local features, quantizing local features into a set of visual words • The method is robust to variances in occlusions, viewpoints, illumination, scale and background • Scale-invariant feature transform (SIFT) features are extracted from representative views of one 3 D object and Bo. W features is generated for a 3 D model representation. Future Directions Many challenges and still require further investigation, including: Large-scale data management Feature extraction Multiview matching Multimodal data Geolocation-based applications