Latest update: Feb 20, 2017
We evaluate results submitted from 5 participants. Each team is identified by the last name of the first author.
In general, the methods proposed by Kanezaki, Chiang, and Truong are based on supervised learning which has considerably higher scores than unsupervised learning methods proposed by Tashiro and Li.
Each table below lists the evaluation results of the test and validation sets.
In the above results, the precision scores of some supervised learning methods appear to be quite low. This is due to the fact that the retrieval results contains CAD models that are have different categories from that of the query object. We also provide results from a slightly different evaluation strategy which only considers the first K retrieved objects, where K is the number of objects in the ground truth class. The scores are as below.