3D volumetric reconstruction from incomplete point clouds still remains one of the fundamental challenges in the process of interpreting the environment, by machines. One solution is to register an a-priori defined shape (i.e. a CAD model) onto the incomplete point cloud of the object. The subsequent challenge that arises is to identify the most similar shape of the object from a shapes database. In this paper, we propose a volumetric reconstruction apparatus that uses so-called Generic Primitive Primitives (GFP) to abstract the large variety of shapes that an object may have. We use a kernel-based deformation technique to fit a GFP to real-world objects, where the kernel is encoded within the layers of a Deep Neural Network (DNN). The objective is to transfer the particularities of the perceived object to the raw volume represented by the generic primitive. We show that GPFNet outperforms competing algorithms on 3D volume reconstruction challenges, by being tested on the ModelNet and KITTI benchmarking datasets. GFPNet is compared with a baseline approach, as well as with state-of-the-art data-driven approaches for volumetric reconstruction.
Implemented and tested on Windows 10 x64, Python 3.5 and Tensorflow 1.12.0rc0
- Clone this repo
git clone git@github.com:RovisLab/GFPNet.git
- Install Python dependencies
pip3 install -r requirements.txt
To train on Kitti Object Detection Dataset: - Download the data and place it in your dataset folder at ~/kitti/
- The folder should look something like the following:
- kitti
- calib
- label_2
- velodyne
- image_2