For this project, we use data from the Waymo Open dataset. The files can be downloaded directly from the website as tar files or from the Google Cloud Bucket as individual tf records.
Bear in mind that you will require to request access to get the buckets in the waymo google cloud.
The data for this project is organized in the following way:
Images/
- All images for the writeup and for any other illustration purposes
build/
- In case you want to run the project in a local computer you should use Dockerfiles inside here
experiments/
- exporter_main_v2.py: to create an inference model
- model_main_tf2.py: to launch training and evaluation of your model
- pretrained-models/: contains the checkpoints of the pretrained models. This folder should be downloaded using the following command:
wget download.tensorflow.org/models/object_detection/tf2/20200711/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.tar.gz
- Exploratory_Data_Analysis.ipynb: In this file you can find all EDA performed such as distribution of the classes.
- Explore_augmentations.ipynb: In this file you can find the training and evaluation of the starter network, tryouts of different augmentation to increase performance and training and evalion steps from them.
- animation.mp4: A video of the network trained from one of the tfrecords
- animation_2.mp4: A second video of the network trained from another of the tfrecords.
- colab_setup.py: This file helps to setup the environment if you chose the docker option. However, this project was done entirely in google colab, so it is not needed.
- create_splits.py: As the name mentions, this file creates the splits between training, validation and testing.
- download_process.py: This file donwloads all data from Waymo's google cloud buckets
- edit_config.py: This file generates new config pipelines to train and evaluate our network
- filenames.txt: Waymo dataset is huge, thus this file lists only the tfrecords that we will use for the scope of this project.
- inference_videp.py: This file creates a video based on a SSD network and a tfrecord
- label_map.pbtxt: This file contains all class labels to map the numeric values from the boxes
- pipeline.config: This is the general pipeline that we use to add more augmentations and anyother changes such as the optimizer
- pipeline_augmentation.config: This file contains out attempt of increasing performance of pipeline.config file through data augmentation
- setup_google_colab.ipynb: This file is key as it contains all the dependencies for this project. This project was entirely developed inside google colab.
- utils.py: All helper functions for the other files above.
- writeup.md: This file contains further details on how the project was developed and all results got from it.
For local setup if you have your own Nvidia GPU, you can use the provided Dockerfile in the build directory or you can run the file setup_google_colab.py to create the environment in google colab.
- In case you choose the docker you can Follow the README therein to create a docker container and install all prerequisites.
- For google colab you should copy the
setup_google_colab.pyand a zip of all the repo to the google colab folder and just call thesetup_google_colab.pyfile to create all the required setup in one cell.
Add the following lines to build the environment for this project. (only works in google colab):
!python colab_setup.py
import os
os.environ['PYTHONPATH'] += ':/content/models/research/:/content/models/research/slim/'
os.environ['PYTHONPATH'] += ":/content/models"
Note: The colab_setup.py file is based on the following reference
The writeup for this project can be found in the following markdown file writeup.md.