A recreation of PoseNet as referenced in this research paper and implemented here. This will be used to determine the distance a visual input is from an object, as well as the input's orientation.
- PyTorch (0.4.0)
- (Dependencies required of this version are implicitly included)
- Singularity (2.5.1)
- The recipes use Ubuntu 16.04 from Docker
- Python (3.6)
- Anaconda (5.1)
- This is not required but this is how we installed our packages into the Singularity container, and is recommended.
- matplotlib (2.2)
- Required because of plotting.
- torchviz (v0.0.1)
- Not required but if you want to visualize the network, you must install this library using pip.
The branches are described in this section.
- Master - Master branch. This branch is protected from pushes from all and merges from outside sources to protect its integrity. If you want to make an edit to the master branch, submit an issue and request for someone within the project to work on it. Merge requests can only be made from Masters and no one can push to it.
- Caffe - The branch that uses Caffe. If it is decided to switch back to using Caffe, this branch is to prevent such a decision from forcing project contributors to start from scratch. This branch likely doesn't work and is protected to preserve its integrity. Caffe is protected to allow merges from only Masters and pushes from no one.
- dev - The development branch that is the precursor to master branch.
To get all avaiable arguments, run python3 main.py --help to get all the arguments of the program.
To visualize the network, run jupyter notebook and open viewPoseNet.ipynb. For an already existing visualization,
open visualize/<version>/viewPoseNet.md.
The pretrained model is the GoogLeNet model provided by hazirbas at GitHub