4.8 SHREC 2021 Object Retrieval Challenge 101 contradicting the assumption of LSBD-VAE that the underlying group action is regular. Nevertheless, our good scores from∆VAE-TL show that there can be merit in combining triplet loss with a VAE-like structure. Table 4.11: Evaluation measures for the retrieval-by-shape challenge. This table shows all the evaluation metrics for all the submitted runs. Highest values are shown in bold. Methods NN FT ST mAP nDCG TL (ours) 0.8159 0.8230 0.8551 0.8409 0.9262 AE-TL (ours) 0.6751 0.4727 0.3578 0.4897 0.8247 VAE-TL (ours) 0.7870 0.7252 0.6895 0.7693 0.9106 ∆VAE-TL (ours) 0.8412 0.8297 0.8357 0.8428 0.9293 LSBD-VAE-TL (ours) 0.7437 0.5385 0.4667 0.5717 0.8482 DNE (run1) 0.7509 0.5813 0.4385 0.6273 0.8704 DNE (run2) 0.8014 0.5933 0.4037 0.6303 0.8746 DRF-L (run1) 0.7256 0.6494 0.5162 0.6780 0.8755 DRF-L (run2) 0.7220 0.4744 0.3493 0.4841 0.8220 MVCNN (run1) 0.7509 0.6164 0.4927 0.6489 0.8794 MVCNN (run2) 0.7545 0.6006 0.4995 0.6350 0.8756 MVCNN (run3) 0.7545 0.8010 0.8860 0.8235 0.9122 MVResNet (run1) 0.1841 0.1430 0.1409 0.1761 0.6281 MeshNN (run1) 0.5054 0.3678 0.2941 0.3673 0.7587 MeshNN (run2) 0.5993 0.6388 0.7358 0.6915 0.8428 MeshNN (run3) 0.7220 0.7871 0.8788 0.8054 0.9013 MeshNN (run4) 0.7906 0.8356 0.9174 0.8518 0.9256 NPC (run1) 0.6751 0.3766 0.3106 0.4019 0.7874 NPC (run2) 0.6787 0.4151 0.3237 0.4274 0.7996 PCI (run1) 0.6751 0.7100 0.7827 0.7496 0.8695 RVN(run1) 0.7942 0.6675 0.5863 0.7084 0.8995 RVN(run2) 0.7726 0.6715 0.5763 0.7074 0.8989 RVN(run3) 0.7762 0.6694 0.6168 0.7086 0.8992 RVN(run4) 0.7798 0.6680 0.6095 0.7069 0.8994 RVN(run5) 0.8087 0.8471 0.9065 0.8604 0.9286 SE3D (run1) 0.7112 0.4380 0.3340 0.4514 0.8068 SE3D (run2) 0.6029 0.4903 0.4126 0.5081 0.8119 SE3D (run3) 0.6354 0.4757 0.4035 0.5065 0.8149 SE3D (run4) 0.5993 0.5090 0.4147 0.5431 0.8205 SE3D (run5) 0.5848 0.5018 0.4358 0.5395 0.8169 Table 4.12 shows the results of the retrieval-by-culture challenge, for our submissions as well as all others. In this case we see that our VAE-TL method scores best amoung our own variants, and also achieves the highest score in one metric overall (and second-highest in the other metrics). Again we see that
RkJQdWJsaXNoZXIy MjY0ODMw