This project hosts the code for our IEEE TCDS paper
An overview of the proposed 3D_DEN model: Initially, three representative views are chosen from a set of multi-view images for a given 3D object.Then, each of them is converted to a single channel (grey-scale) image and later merged to form a 3-channel image. Now, this image is fed to a pre-trained network, and the extracted features are flattened. Finally, we attach two DEN layers to the model which give the output.
- Python 3.6
- Kindly create a virtual environment using requirements.txt file to run the code
- Note: For Offline Evaluation using GridSearch, use Tensorflow and Tensorboard version: 2.3.0.
Latest version available on arXiv (March 2021) | Video | Report
Please adequately refer to the paper any time this code is being used. If you do publish a paper where 3D_DEN helped your research, we encourage you to cite the following paper in your publications:
@ARTICLE{jain-3dden-2021,
author={Jain, Sudhakaran and Kasaei, Hamidreza},
journal={IEEE Transactions on Cognitive and Developmental Systems},
title={3D_DEN: Open-ended 3D Object Recognition using Dynamically Expandable Networks},
year={2021},
doi={10.1109/TCDS.2021.3075143}
}
Sudhakaran Jain and Hamidreza Kasaei
Work done while at RUG.