Section: Bilateral Contracts and Grants with Industry

Bilateral Contracts with Industry

CIFRE contract with Technicolor on light fields editing

Participants : Christine Guillemot, Matthieu Hog.

• Title  : Light fields editing

• Research axis : 7.1.1

• Partners : Technicolor (N. Sabater), Inria-Rennes.

• Funding : Technicolor, ANRT.

• Period : Oct.2015-Sept.2018.

Editing is quite common with classical imaging. Now, if we want light-field cameras to be in the future as common as traditional cameras, this functionality should also be enabled with light-fields. The goal of the PhD thesis is to develop methods for light-field editing, and in 2018 we have extended our concept of super-rays initially introduced for static light fields to video light fields (see Section 7.1.1). Super-rays group rays within and across views, emitted by the same set of 3D points in the space. A method for dynamic tracking of super-rays with scene flow estimation has been developed. We have further explored a novel way, using recurrent neural networks and in particular long short term memory (LSTM) networks, to solve the problem of view synthesis (see Section 7.3.1).

CIFRE contract with Technicolor on light fields compressed representation

Participants : Christine Guillemot, Fatma Hawary.

• Title  : Light fields compressed representation

• Research axis : 7.2.2

• Partners : Technicolor (G. Boisson), Inria-Rennes.

• Funding : Technicolor, ANRT.

• Period : Feb.2016-Jan.2019.

The goal of this PhD thesis is to study reconstruction algorithms from compressed measurements. The goal is to apply these algorithms to scalable compression of light fields. Methods of sparse light field recovery have been developed, based on the assumption of sparsity in the Fourier domain, and using orthogonality constraint in the Fourier transform domain. The method has been further improved by introducing a refinement of the basis functions with non integer frequencies.

CIFRE contract with Technicolor on cloud-based image compression

Participants : Jean Begaint, Christine Guillemot.

• Title  : Cloud-based image compression

• Research axis : 7.2.6

• Partners : Technicolor (Ph. Guillotel, F. Galpin), Inria-Rennes.

• Funding : Technicolor, ANRT.

• Period : Nov.2015-Oct.2018.

The goal of this Cifre contract is to develop a novel image compression scheme exploiting similarity between images in a cloud. A region-based geometric and photometric alignment algorithm has been developed and validated for still image compression with an inter-coding set-up using similar images in the cloud as reference frames. This model has been further validated in the context of temporal prediction in a video compression scheme (see Section 7.2.6). Neural network based frame interpolation techniques have also been investigated, showing promising performance gains compared to the state of the art.

DGA contract on deep learning for image compression

Participants : Thierry Dumas, Christine Guillemot, Aline Roumy.

• Title  : Deep learning for image compression

• Research axis : 7.2.5

• Partners: Inria-Rennes (Sirocco team)

• Funding: DGA/Ministry of defense

• Period : Oct.2015-Sept.2018.

This project funded by the DGA/Ministry of Defense concerns the PhD thesis of T. Dumas. The goal was to study deep learning architectures for image compression. Autoencoders have been studied to jointly learn transforms and quantizers with rate-distortion optimization criteria. A set of neural network architectures called Prediction Neural Networks Set (PNNS), based on both fully-connected and convolutional neural networks, has also been developed for intra image prediction (see Section 7.2.5).