Home
Uni-Logo
 

Oberseminar Bildverarbeitung, Computersehen und Computergraphik

Prof. Dr.-Ing. Thomas Brox, Prof. Dr.-Ing Matthias Teschner, apl. Prof. Dr. Olaf Ronneberger

Nächste Termine:
11.07.2017

Vorankündigung (Aushang):

Dienstag, 11.07.2017, 14:00 h
Geb. 52, SR 01-033

Herr Denis Tananaev
Institut für Informatik, Universität Freiburg

berichtet über die Ergebnisse seiner Masterarbeit:

Structure from Motion with Recurrent Neural Networks



In this work we develop recurrent convolution neural networks for depth prediction and position estimation given a video sequence from a monocular camera. We compare the performance of 2D and 3D residual convolutional networks with networks that use convolutional LSTM with tanh or leaky ReLU activations. We show the importance of layer normalization for designing non-saturating recurrent networks which are able to process any length of input sequences. Our experiments show that the recurrent networks allow to accumulate the information from all previously observed frames and thereby improve the quality of predictions. Our approach outperforms frame to frame processing by the 2D and 3D convolutional feed-forward networks.



Bitte beachten Sie die Zeit und den Ort!

Interessenten sind herzlich eingeladen. Weitere Informationen bei: Prof. Dr.-Ing. Thomas Brox, Tel: 0761/203-8261



Vorankündigung (Aushang):

Dienstag, 11.07.2017, 15:00 h
Geb. 52, SR 01-033

Herr Osama Makansi
Institut für Informatik, Universität Freiburg

berichtet über die Ergebnisse seiner Masterarbeit:

End-to-End Learning of Video Super-Resolution with Motion Compensation



The task of reconstructing higher resolution video frames given the low resolution ones has been studied for a long time. Recently deep learning techniques started to achieve promising results in terms of accuracy and run-time, solving many computer vision tasks including the video super-resolution task. The estimated motion between the input frames plays an important role in predicting the high resolution frame. Using this motion is utilized by warping the input frames to compensate for that motion. Many of the recent approaches used precomputed flows to warp the frames and then trained a network for predicting the output high resolution frame. In this report, we analyze some recent approaches and propose an end-to-end network that can estimate jointly the motion and the high resolution frames. On the other hand, we test many architectures for single image super-resolution and find that a simple architecture consisting of only few layers can achieve good accuracy. Moreover, we provide a new operation combining upsampling and backward warping (named as JUBW) on the input frames in order to compensate for the estimated motion. This method warps the neighboring low resolution frames toward the center frame and fuses their details into the high resolution grid without any interpolation. Finally, we conduct experiments where we compare our architecture including the proposed warping method to other methods. Our testing results are superior to the current state-of-the-art approaches.



Bitte beachten Sie die Zeit und den Ort!

Interessenten sind herzlich eingeladen. Weitere Informationen bei: Prof. Dr.-Ing. Thomas Brox, Tel: 0761/203-8261