M5 Project: Scene Understanding for Autonomous Vehicles

The goal of this project is to learn the basic concepts and techniques to build deep neural networks to detect, segment and recognize specific objects. These techniques will be applied to environment perception for autonomous vehicles, whereby the classes of interest with regard the three tasks will be pedestrians, vehicles, road, roadside, etc.

Team 6 members

• Alba Herrera (albaherrerapalacio at gmail dot com)
• Jorge López (jorgelopezfueyo at gmail dot com)
• Oscar Mañas (oscmansan at gmail dot com)
• Pablo Rodríguez (pablorodriper at gmail dot com)

Running instructions

pip3 install -r requirements.txt
python3 src/main.py --config_file CONFIG_FILE

Directory structure

.
├── config              # framework configurations
├── devkit_kitti_txt
├── docs                # summaries of nn systems
│   ├── resnet.md
│   └── vgg.md
├── fonts
├── jobs                # jobs to schedule in the SLURM cluster
├── README.md
├── requirements.txt    # python dependencies
└── src
    ├── config
    ├── dataloader
    ├── main.py
    ├── metrics
    ├── models
    ├── tasks
    └── utils

Weekly Deliverables

Week 2 - Object Recognition

• Summary
• Slides
• Model weights:
  • OscarNet on BelgiumTSC

To run the framework with the weights above, execute:

python3 src/main.py --config_file config/oscarnet_tt100k_pretrained.yml

Weeks 3,4 - Image Semantic Segmentation

• Summary
• Slides
• Model weights:
  • PSPNet50 on CamVid
  • PSPNet101 on CamVid

Weeks 5,6 - Object Detection Segmentation

• Repository
• Summary
• Slides
• Model weights:
  • RetinaNet on TT100K

Report

Overleaf link

Related publications

Object Recognition

VGG (CVPR 2014):

• Original paper
• Summary

ResNet (2015):

• Original paper
• Summary

Image Semantic Segmentation

FCN (CVPR 2015):

• Original paper
• Summary

PSPNet (CVPR 2017):

• Original paper
• Summary

Object detection

Faster R-CNN (NIPS 2015):

• Original paper
• Summary

RetinaNet (ICCV 2017):

• Original paper
• Summary