Single images 3D reconstruction by a binary classifier

Abstract

Intelligent systems demand interaction with a variety of complex environments. For example, a robot might need to interact with complicated geometric structures in an environment. Accurate geometric reasoning is required to define the objects navigating the scene properly. 3D reconstruction is a complex problem that requires massive amounts of images. The paper proposes producing intelligent systems for 3D reconstruction from single 2D images. Propose a learnable reconstruction context that uses features to realize the synthesis. Proposed methods produce encoding feature lable input to classification, pulling out that information to make better decisions. Binary Classifier Neural Network (BCNN) classifies whether a point is inside or outside the object. The reconstruction system models an object 3D structure and learns feature filter parameters. The geometry and the corresponding features are implicitly updated based on the loss function. The training doesn’t require compressed supervision to visualize the task of reconstructed shapes and texture transfer. A point-set network flow results in BCNN having a comparable low memory footprint and is not restricted to specific classes for which templates are available. Accuracy measurements show that the model can extend the occupancy encoder by the generative model, which doesn’t request an image condition but can be trained unconditionally. The time required to train the model will have more neurons and weight parameters overfitting.

Keywords

• intelligent systems
• 3d reconstruction
• features filter
• convolution neural networks
• binary classifier neural network (bcnn)