Scalable 3D partial object retrieval with spatial relationships
vendredi 14 mars 2014, 11h00 - 12h00
This work presents an approach for 3D object retrieval, dedicated to partial shape retrieval in large datasets. Sets of 3D words, i.e. quantized 3D descriptors as in the Bag-of-Words representation, are employed, based on the extraction of 3D Harris points and on a local description involving local Fourier descriptors. By adding Î”-TSR, a triangular spatial information between words, the richness and robustness of this representation is reinforced. The approach is invariant to different geometrical transformations of 3D shape like translation, rotation, scale and robust to shape resolution changes. A dedicated disk-based indexing structure is also employed to make Î”-TSR effective on large-scale datasets. We have evaluated it in terms of quality of retrieval and scalability, facing several state-of-the-art methods and on public 3D benchmarks involving different contents and degrees of complexity.