Abstract
Sketch-based 3D model retrieval focus on retrieving relevant 3D models using sketch(es) as input. It is attractive to use sketch in category-level 3D retrieval because sketches by hand provide an easy way to input, yet they are detailed enough to specify shapes. Compared to category-level sketch-based 3D retrieval, instance-level retrieval gives user more pragmatic retrieval results. The users draw the sketch, and the system will yield the exact 3D model that is desired. However, as the number of 3D models is limited in the dataset, there might not exist the exact 3D model that the user wants. But, we could get a similar alternative by assembling suitable 3D parts in the current dataset. This underlines the need for part-level sketch-based 3D retrieval. In this task, one pre- requirement is that sketches need to be segmented by parts. This emphasises the need for practical semantic sketch segmentation.
In this thesis, we explore the problem of sketch-based 3D retrieval in a coarse-to-fine manner, i.e., from category-level to instance-level to part-level retrieval. And show that sketch is especially powerful when communicating with the high-dimensional 3D world, where currently humans still have difficulty in interpreting its visuals and understand- ing how interaction occurs. Besides, to facilitate the part-level sketch-based 3D shape retrieval, we address the problem of personalised sketch segmentation for the first time.
For all the aforementioned tasks, we perform rigorous experiments on various publicly available datasets. The quantitative and qualitative results are included to demonstrate the effectiveness of our proposed solutions.