The reconstruction of 3D objects from 2D images is a growing topic in machine learning, and the problem proves to be challenging as we are faced with computations that grow in cubic form as we attempt to process larger inputs. 3D- ReConstnet is a point-cloud-based 3D reconstruction algorithm that aims to eliminate the concerns with voxel-based occupancy grid computational complexity. In this project, we want to explore the architecture of 3D-ReConstnet and recreate the network by hand. After recreating the network, we want to explore improvements and optimizations, so that we can improve the resulting point-cloud generation.
implementation is as followed:
- Obtain 2D images from the ShapeNet Dataset [3] and extract its necessary features to a feature vector using the ResNet-50 Network. Different combinations of input to the network will be experimented with to obtain different outputs.
- Obtain the mean and standard deviation of the feature vector and obtain the Gaussian probabilistic vector to encode the results.
- Perform decoding of the Probabilistic Vector by imple- menting a multi-layer perception.
- Experiment with different loss functions and training parameters to improve 3D results of output image.
