Vehicle Detection for Cyclist Safety Training
This repository contains code to facilitate the training of a deep neural network for vehicle detection.
In order to begin training, you must have a directory structure similar to the following:
- training
- data
- train.record
- val.record
- models
- model
- pipeline.config
- model
- data
An example config for ssdlite_mobilenet_v2 is given in the home directory: ssdlite_mobilenet_v2_vehicle_detection.config. The commands to use to start training are given in train.sh which should be used as a guide (but may not work by simply running . train.sh.)
In order to create data, generate_tfrecord.py and create_coco_tf_record.py are provided. The former is made to use data from the udacity self-driving car dataset while the latter uses the relevent classes from mscoco (bicycle, car, motorcycle, bus, truck). In order to run the script with the correct inputs, coco.sh can be used as a guide.
In order to ensure that the correct images have been written to the tfrecord, view_images.ipynb can be used to view the images in a tfrecord.
Once training is complete, freeze_model.sh can be used to freeze the model for deployment.