In recent times, the researchers from deep learning community started focusing on 3D modeling for broader adoption. The algorithms are being implemented at a fast pace to work with 3D data. The main goal is to reconstruct and generate 3D models that can help save time in modeling 3D data in various industry domains like robotics, computer vision, and virtual reality, etc., However, the quality of 3D generation is not attained to the required level. In this thesis, we present an implicit generative framework that merges implicit representation and normalizing flow techniques to generate novel shapes. Implicit representation excels at intricate 3D shape learning while normalizing flow-based generation facilitates novel 3D shape generation offering tractable log-likelihoods and efficient sampling. This work further analyzes the latent spaces in neural implicit representation models, aiming to identify the optimal shape learning models. This enhances the ability to generate novel shapes by learning the optimal shape encoding. By capturing 3D shapes as dense point clouds, our approach advances generative 3D modeling. We conduct our experimentation on the ’cars’ object category from the ShapeNet dataset for its intricate internal details and complexity. To quantitatively assess the efficacy of our 3D reconstruction models, we employ the chamfer distance metric. Additionally, we present qualitative outcomes of our generation approach to provide a comprehensive view of our model performance.
The NDF.yml file contains all necessary python dependencies for the project. To conveniently install them automatically with anaconda you can use:
conda env create -f NDF.yml
conda activate NDF
To perform the experiments ShapeNetCore v2 dataset cars foler named as 02958343 is downloaded.
-
Dataprocessing is performed with:
python ../dataprocessing/preprocess.py --config ../configs/shapenet_cars.txt -
The split of training, test or validation split data using:
python ../dataprocessing/create_split.py --config ../configs/shapenet_cars.txt -
Train the Real NVP model with the following command:
python ../train_rnvp.py --config ../configs/shapenet_cars.txt -
Visualization of graphs:
tensorboard --logdir ../models/experiments/shapenet_cars/summary -
Generate 3D car models:
python ../generate.py --config ../configs/shapenet_cars.txt
Checkpoints can be found in ../models/experiments/shapenet_cars for NDF model to freeze the checkpoint.
For running the python files the ../slurm_scripts folder contains shell scripts.
Note: Checkpoints have to be changed to the format checkpoint:h:m_s_.tar from checkpoint_h_m_s_.tar before running the experiments.