This project is done in the framework of the Master in Computer Vision by Universitat Autònoma de Barcelona (UAB) in the module M5: Visual recognition.

Abstract: "In this project we employ state-of-the-art Deep Learning algorithms for recognition, detection and segmentation tasks which are key to create a system which understands traffic scenes correctly and makes the right decision while autonomously driving".

For more information, check our report on Overleaf or our presentations on Google Drive. You can find summaries for some Image Classification papers on the file summaries. Additionally, for a quick look at what we have done in the project so far, head over to the 'The progress at a glance section'.

• Getting Started
• How to run it
• The progress at a glance
  • Object Recognition
  • Object Detection
• Authors
• License

The software need the following software to run:

• Python 2.7
• pip
• Pipenv

If you are using any Linux distribution is most probable you already have a running version of Python installed.

As an example, if you are using an Ubuntu you can install pip and pipenv using these commands:

apt-get install python-pip
pip install pipenv

To install the virtual environment you only have to run pipenv from project's root directory:

pipenv install

It is recommended to add the following environment vars to your .bashrc:

export PIPENV_VENV_IN_PROJECT=1
export PIPENV_IGNORE_VIRTUALENVS=1
export PIPENV_MAX_DEPTH=1

You can find their explanation as well as more environment variables in configuration with environment variables.

This project is using Keras as a high-level neural networks API running on top of Tensorflow library.

cd code
CUDA_VISIBLE_DEVICES=0 python train.py -c config/tt100k_classif.py -e test -l /home/master/tmp -s /data/module5/

python code/train.py -c code/config/tt100k_classif.py -e test -l tmp -s data

You can find some the weight from our experiements in this GoogleDrive.

Two weeks have been devoted to the study of state-of-the-art architectures for object recognition. In particular, we have evaluated the vainilla VGG16 architecture weights to recognise traffic signals from the TT100K dataset. We compared the performance between cropped and resized images as well as the adaption of the network to another domain (BelgiumTS). Moreover, we have trained the VGG16 both from scratch and with pre-trained weights from ImageNet on the KITTI dataset and analysed their results. Finally, we have implemented the Squeeze and Excitation Network (added to models and compared the vainilla ResNet50 results with its SE counterpart, both with fine-tuning and from scratch. Each dataset has been analysed to help us draw meaningful conclusions from the results obtained. Check our report and presentation for more details.

Two weeks have been devoted to the study and implementation of state-of-the-art architectures for object detection. We have tested the vainilla YOLO (v1) architecture with both TT100K (for detection, instead of using crops containing only signals) and Udacity datasets and analysed overffiting and unbalancing problems we encountered. We have proposed possible solutions and implemented some, like data augmentation and re-splitting the partition given. We also worked hard in implementing an SSD but we could not finished it. We can see the loss decreasing and getting to a plateau during the training but we have no bounding boxes to assess the detections. Check our report and presentation for more details.

• Ferran Pérez - dev - fperezgamonal - contact
• Joan Francesc Serracant - dev - fserracant - contact
• Jonatan Poveda - dev - jonpoveda - contact
• Martí Cobos - dev - marticobos - contact

This project is licensed under the GPLv3 License - see the LICENSE.md file for details