facebook tracking

Master Thesis - Importance sampling in object detection training

Curious about how Deep Learning can be used for the development of self-driving vehicles? Explore more in this master thesis!

Background 

Efficient and robust detection of objects under a broad range of possible circumstances is a central challenge in the development of autonomous vehicles. The object detector at Zenseact is a deep neural network trained on a large number of frames captured by on-car cameras. The resolution of the input images to the object detection network is necessarily large to ensure a sufficient performance. The combination of high resolution frames and a large data volume results in an immense computational cost for training the neural network. Currently the neural network is trained with minibatch gradient descent in a number of iterations over the entire training data set. During the first few iterations the training of the network converges relatively quickly and most frames produce substantial gradients. At later stages of training, the network training slows down significantly. To a large extent this is caused by the fact that the network is already able to produce correct predictions for the bulk of the frames in the training set, resulting in tiny gradient updates.

A fraction of frames, corresponding to objects and circumstances more difficult to handle for the network, will keep producing relatively large gradients. Such frames significantly further the network training. Oversampling the frames that have a high impact on the learning of the network might be a powerful way of decreasing the training time of the object detection network, while achieving a similar network performance in the end. This would have a host of benefits: Issues could be addressed more quickly in new version of the network, more experiments could be carried out for improving the network performance, and as the data volume keeps growing larger and larger the training time would stay manageable.

We propose a thesis in which the students would systematically investigate the possibility of improving the training times of the object detection network by means of importance sampling strategies.

Project Description 

In this master thesis project, you will

  • Develop methods for clever sampling of frames during neural network training.
  • Test how well these methods manage to speed up the training and improve the final performance.
  • Implement promising methods in Zenseact's training framework.

Qualifications 

We are looking for 2 students with an interest in deep learning for autonomous driving. The following skills would be highly valuable:

  • Python programming
  • Machine learning
  • Reading scientific papers
  • Handling large datasets

 

Further information

Please send in individual applications with CV, motivational letter and grade transcripts. 

Planned start: January 2022, with some flexibility.

Final application date: 15 of November 2021, but we will screen candidates continuously, so please submit your application as soon as possible.

Duration: 30 ECTS 

For questions regarding the project, please contact: willem.verbeke@zenseact.com and erik.werner@zenseact.com

Additional information

  • Remote status

    Flexible remote

Or, know someone who would be a perfect fit? Let them know!

Gothenburg, Sweden

Lindholmspiren 2
417 56 Göteborg Directions View page

Making safe and intelligent mobility real.

At Zenseact, we lead the global movement of crafting tomorrow's mobility with the software platform of choice. Our mission is to “Make safe and intelligent mobility real, for everyone, everywhere”. This statement marks our conviction and dedication to bring autonomous driving out on the streets for real and is at the center of everything we do.

We could not dream of achieving this without our great teams of very talented people. We are on this journey together and our agile way of working is reflected throughout our entire organization; it is part of our culture and how we work, develop and grow together.

Teamtailor

Applicant tracking system by Teamtailor