4.8 SHREC 2021 Object Retrieval Challenge 95 Figure 4.15: Sample objects for every class of the retrieval-by-culture dataset. Image from Sipiran et al. (2021). Both datasets are split into a training set (70% of the dataset, used for training and selecting hyperparameters) and a test set (30% of the dataset, only used for evaluation after the challenge was finished). For both challenges, the classes are not balanced, creating an extra challenging but realistic scenario for 3D retrieval algorithms. For more information about the challenge, see Sipiran et al. (2021). 4.8.2 Our Methodology For our submissions to this retrieval challenge, we first obtain 2D images rendered from the 3D models, such that we can use image-based representation learning methods combined with metric learning to accommodate retrieval. Specifically, we use various autoencoder-based methods to learn low-dimensional representations of the 3D objects (via the rendered 2D images), where we use a triplet loss to encourage similar objects to have similar representations. Note that similarity here is defined by the challenge, based on shape or culture, respectively. In particular, this is a suitable setting for our own LSBD-VAE, since we can obtain several images of the same object under various orientations (i.e. SO(2) symmetries).
RkJQdWJsaXNoZXIy MjY0ODMw