The MTL model has been successfully implemented, leveraging the benchmarked NITUAVCorridorV1 dataset for training and evaluation. I utilized the pre-trained deep learning model DenseNet161.
- DenseNet architecture primarily consists of two types: Dense blocks and Transition Layers that handle convolution and pooling operations.
- Within these transition layers, I defined task-specific branches to facilitate task-specific training for the model.
- This approach allows for efficient utilization of the pre-trained DenseNet model while adapting it to the specific requirements of our project.
- Torch
- Torchvision
- pandas
- numpy
- matplotlib
- scikit-image
- virtualenv
You can install the NITUAVCorridorV1 dataset here:
http://www.nitrkl.ac.in/docs/CS/Database/Windows/NitrUAVCorridorV1.zip
Unzip the file and can access the dataset to run this model.
Firstly try to create a virtual environment for better code base establishment.
python -m virtualenv <your_env_name>
Once you create, u can use this command to activate it.
source <your_env_name>/bin/activate
Run the main.py to get the model running.
Here are few plots of the model after 100 epochs of training and testing with the 900 evaluation images:
-
R. P. Padhy, S. Ahmad, S. Verma, S. Bakshi, and P. K. Sa, “Localization of un- manned aerial vehicles in corridor environments using deep learning,” in 2020 25th International Conference on Pattern Recognition (ICPR), 2021, pp. 9423– 9428
-
Y. Zhang and Q. Yang, “An overview of multi-task learning,” National Science Review, vol. 5, no. 1, pp. 30–43, 09 2017. [Online]. Available: https://doi.org/10.1093/nsr/nwx105
-
S. Ruder, “An overview of multi-task learning in deep neural networks,” arXiv preprint arXiv:1706.05098, 2017
-
Amaarora. (2020, August) Densenets. [Online]. Available: https://amaarora. github.io/posts/2020-08-02-densenets.html