Scene understanding in computer vision involves many tasks to make it an affordable problem. Among them, there is the problem of removing, or at least reducing, the effects of lighting in the scene. An specific example is the automatic guiding assistance in vehicle driving. Computational methods, that have to cope with understanding what is going on in a given scene, have to deal with shadows, specularities, changes on color and intensity of light, etc. All these effects, which are related to the physical image formation, are specially problematic in the field of computer vision systems for transportation. Next figure is a graphical example of these scenarios.
The literature contains partial solutions that mainly use known physical laws to understand what is a shadow, a highlight, etc. These methods introduce this knowledge as priors in a classical classification framework, where other cues might not be taken into account. But, which are the relevant cues for a given problem?.
On the last years a sound methodology -Convolutional Neural Networks (CNN)- has been applied to extract the relevant features to solve a given problem. CNN do not use any knowledge about the problem apart from examples and counterexamples.
Humans can learn by examples but, when provided with basic rules their learning process is shorter and more efficient. Right now, there is not a way of introducing this human behaviour in CNN. The aim of the PhD proposal is to investigate which is the ideal framework in machine learning framework, including CNN and MRF to adapt known physical laws to automatic knowledge extraction. the goal is to drive the learning processes by means of physical model constraints, allowing to reduce the amount of data needed to improve state of the art results.. The work will be focused on learning to classify which parts of an image are cast shadows, self shadows, and specularities.
One of the problems when working with Neural Networks is to have a reliable ground truth dataset. Part of the work will be focused on setting up a dataset with images either real or synthetic to provide tenough data onto the machine learning models.
The ideal candidate for the position must have:
- B.Sc. and M.Sc. in Computer Science, Electrical Engineering, Computer Engineering, Telecommunications, or a related field (concluded by the time you join).
- A background in computer vision and/or computational colour science.
- Excellent academic transcripts.
- Proficiency in spoken and written English (please note: Spanish is not required).
- Good communication skills.
- High level of motivation and a strong interest in a scientific career.
The PhD will be developed at the Computer Vision Center (http://www.cvc.uab.cat) in the ‘Colour in Context’ research grouphttp://cic.uab.cat/.
How to apply
This call is part of a wider call from the Computer Science Department of the Autonomous University of Barcelona. The period for presenting applications is from October, 30th to November, 7th, 2015.
Candidates must accredit that they hold a university qualification, EHEA university qualification of at least 300 ECTS credits, or a university-awarded Master degree or equivalent, or be able to accredit this at the time of signing the contract. Candidates cannot already hold a doctoral degree. The forecast date of incorporation is January, 11th 2016.
Interested candidates must send an e-mail to email@example.com with an expression of interest specifying the above profile, and attaching the following documentation:
- A copy of the DNI or passport
- A certificate of the academic record
- A copy of the official certificate of the university degree
Further details on this call can be found (the deadline in the linked document is from a previous call, the conditions are the same)
The duration of each position is of 3 years (36 months). The total annual amount of the grant is €14.391,63 gross, payable in 12 monthly installments that include the pro-rated bonus payments. Grants additionally include the cost of public registration fees for official doctorate programs and supervision of the thesis.
The profile and contact info for this specific position is
Computer Vision – Scene understanding (contact: Ramon Baldrich, firstname.lastname@example.org)