A Comparison of Text and Shape Matching for Retrieval of Online 3D Models with statistical significance testing

Because of recent advances in graphics hardand software, both the production and use of 3D models are increasing at a rapid pace. As a result, a large number of 3D models have become available on the web, and new research is being done on 3D model retrieval methods. Query and retrieval can be done solely based on associated text, as in image retrieval, for example (e.g. Google Image Search [15] and [40, 51]). Other research focuses on shape-based retrieval, using methods that measure shape similarity between 3D models (e.g., [14]). The goal of our work is to take current textand shape-based matching methods, see which ones perform best, and compare those. We compared four text matching methods and four shape matching methods, by running classification tests using a large database of 3D models downloaded from the web [50]. We also investigated several simple methods to combine the scoring results of text and shape matching. Additionally, this report uses statistical tests for significance testing of differences in precision/recall results. We found that shape matching outperforms text matching in all our experiments. The main reason is that publishers of online 3D models simply do not provide enough descriptive text of sufficient quality: 3D models generally appear in lists on web pages, annotated only with cryptic filenames or thumbnail images. Of the available text sources of a 3D model, we found that the text inside the model file was the most useful for classification. Also, adding synonyms and hypernyms (category names) of model filenames using WordNet [31] improved classification results. Combining the results of text and shape matching further improved performance. The results of this work provide added incentive to continue research in shape-based retrieval methods for 3D models, as well as retrieval based on other attributes.