ITS 3D Camera-Only Detection

Detection of cars in the Waymo-Open-Dataset
Report Bug

Table of Contents

1. About The Project
   • Build With
2. Getting Started
3. Important Files and Folders
   • Main file
   • Training
   • Results
4. Contact

About The Project

This project was conducted as part of the course "KV Special Topics in AI - ITS" in WS22.
The primary goal was to look at the Waymo dataset, more specifically the 3D camera-only detection
part of the Motion dataset. After downloading and processing the dataset we tried 2 different models.
We programmed a CNN ourselves as a model and then continued to work with the pre-built UNET.

The UNET was developed by Olaf Ronneberger et al. for Bio Medical Image Segmentation.
The model is an end-to-end fully convolutional network (FCN), i.e. it only contains Convolutional layers and
does not contain any Dense layer because of which it can accept image of any size.

Built With

• PyCharm
• Jupyter
• Github Desktop
• Vim
• GCloud

Getting Started

For this project, only the training dataset was used for space reasons, which is after all also approx. 800GB in size.
The downloaded records are then (or during) selected with the extractor, compressed, processed
and saved as tensors in a pickle file.

Installation

1. Clone the repo

   git clone https://github.com/TristanBandat/ITS_3D_Camera_only_detection.git

2. Install dependencies
   

   The fastest way to install the necassary dependencies is via conda:
   

   conda install -f environment.yml

3. Download dataset from here .
   For downloading the 1GB big tfrecords use the following command:
   

   gcloud storage cp "[FILE]" "[FILE]" ...

4. Select and compress data to a pickle file using the Extractor.
   

   python extractor.py

5. Now one can proceed with the notebook or the main file.

Important files and folders

Main file

main.py
The entry point is the main.py file. Here one can find all the different hyperparameters and available models to train.
Furthermore the path to the dataset and the final models is also chosen here.
If all the necessary packages are installed one can simply run the file and the training starts.

Training

train.py
Here one can find the whole project structure. Further details and explanations are contained in the notebook.

Results

In the results/ folder the final model and the plots can be found. With the help of tensorboard one can also
view additional plots like train/validation loss or the gradients.

Contact

Tristan Bandat - tristan.bandat@gmail.com
Philipp Meingaßner - meingassner.p@gmail.com
Jakob Eggl
Florian Hitzler