3D Patch-Based Sparse Learning for Style Feature Extraction

Abstract

How to extract features of different shapes for style similarity evaluation is a very challenging research topic. Different from methods based on predefined style templates, we perform sparse learning and triplet embedding directly from the 3D shape local descriptors. The proposed method can adaptively generate element-level style features of different 3D patches for style similarity evaluation. The proposed algorithm mainly consists of the following steps. First, given a heterogeneous 3D shape collection, we learn mid-level style features by Siamese networks based on curvature-guided sampling directly on 3D shapes, which is a view-independent and part-independent approach. Second, we perform sparse coding to reduce the dimension of mid-level local features and achieve more compact and discriminative style features, named element-level features. Finally, we enhance our element-level style features by triplet embedding learning over crowdsourced triplets, which maps similar examples close to each other and dissimilar examples farther apart. The proposed algorithm can aggregate different local features into a global style feature representation and shows satisfactory discriminability. Our method can avoid the loss of style features from the 3D model to the views and can learn more discriminative style features by sparse learning. The experimental results show that our sparse style feature representation significantly improves the accuracy compared with recent state-of-the-art algorithms. INDEX TERMS 3D shape analysis, style feature, sparse learning, 3D patch.

UVStyle-Net: Unsupervised Few-shot Learning of 3D Style Similarity Measure for B-Reps

“…Geometric Style Similarity. Existing geometric style similarity learning methods are typically trained in a supervised setting, requiring a set of hand-labeled triplets (A, B, C) in which the pair A and B are believed to be closer in style than A and C [24,25,22,27,28,30]. To account for style subjectivity, examples are labeled through crowd-sourcing methods and thus result in a generally accepted definition for style.…”

“…Rendering 3D solids into 2D (even with multiple views) is problematic since stylistic features can be lost or occluded and selecting the best views without making assumptions on the orientations of the data is non-trivial. Pan et al [28] overcome this using curvature-guided sampling directly from the solids to generate element-level style features which are then aggregated to global style representations using a triplet network.…”

“…While existing methods use hand-crafted features [25,24] or crowd-sourcing [22,27,30,28] to pre-define and measure geometric style, we propose a user-defined fewshot style metric learning method that leverages the range of style signals available in the activations of a pre-trained 3D object encoder through second order statistics (Gram matrices). The relative importance of each layer's Gram matrix is then learnt through selection of just a few examples of what style means to an end user (see Figure 1).…”