This repository contains the codes used for training LiDAR Cone Classifier. The classifier performs classification on the images generated by img_node of the Ouster OS1-64 LiDAR. The LiDAR image crops have their intensity value capped at 1000 and rescaled to [0, 255], then they are exported as 32 x 32 grayscale images.
The list below contains some of the key dependencies required.
imgaug=0.4.0
pytorch=1.5.1
torchvision=0.6.1
pillow=7.2.0
matplotlib=3.3.0
numpy=1.19.1
It is assumed that the use has the folders setup as shown below.
├── data
│ ├── blue [578 entries exceeds filelimit, not opening dir]
│ └── yellow [500 entries exceeds filelimit, not opening dir]
├── README.md
└── train.ipynb
Currently all training and inference operations take place within the train.ipynb notebook.
- Implement basic pytorch image classifier
- GPU checking and GPU training
- Document dependencies and packages required to train the classifier
- Export trained model to ONNX, so that it can be further optimised for TensorRT deployment
- Investigate how to extend this to classify
oragne traffic cones -
pytorch-lightningmodule to improve reproducibility and reduce boiler plate