Point2Cyl: Reverse Engineering 3D Objects from Point Clouds to Extrusion Cylinders

SECAD-Net: Self-Supervised CAD Reconstruction by Learning Sketch-Extrude Operations

“…The sketches in each sketch plane depict the shape contained within the sketch box. Inspired by the recent neural implicit shapes [2,46], we encode the shape of each sketch into a sketch latent space. To this end, we first project the 3D sampling points w.r.t the corresponding occupancy value onto the sketch plane along the axis e i .…”

“…While achieving high-quality reconstruction, CSG tends to combine a large number of shape primitives that are not as flexible as the extrusions of 2D sketches and are also not easily user edited to control the final geometry. Motivated by modern design tools, supervised methods are proposed [22,46] utilizing the sketch-extrude procedural models and learning 2D sketches that can be extruded to 3D shapes. In contrast to their reliance on 2D labels, SECAD-Net is trained in a self-supervised manner.…”

“…We thoroughly compare our method with two types of primitive-based CAD reconstruction methods that output editable CAD models, including two CSG-like methods (i.e., UCSG-Net [19], CSG-Stump [33]) and two cylinder decomposition counterpart (i.e., point2Cyl [46]), Extru-deNet [34]. For each method, we adopt the implementation provided by the corresponding authors, and use the same training strategy for training and fine-tuning.…”

Recognizing geometric primitives in 3D point clouds of mechanical CAD objects

“…The third class gathers all the other clustering techniques, and can be classified into three main types: primitive-driven region growing, e.g., [13]; automatic clustering and Lloyd-based algorithms, e.g., [14]; primitive-oblivious segmentation, e.g., [15]. Finally, with the growing popularity of deep learning techniques, supervised fitting methods have been proposed even for multi-class primitives [16,17]. The reader is referred to [10] for a comprehensive historical taxonomy of methods for simple primitive detection, which is beyond the scope of this paper.…”

A Review of Deep Learning-Powered Mesh Reconstruction Methods

“…Point2Cyl [151] all reconstruct parametric surfaces from point clouds. They first use a neural network to extract per-point features from the input point clouds, and then apply a clustering module to segment the point cloud into patches belonging to different primitives, and finally classify the primitive type and regress the primitive parameters for each patch, as shown in Figure 2.6 (a).…”