Direkt zum Inhalt springen
Computer Vision Group
TUM Department of Informatics
Technical University of Munich

Technical University of Munich

Menu

Links

Informatik IX
Chair of Computer Vision & Artificial Intelligence

Boltzmannstrasse 3
85748 Garching info@vision.in.tum.de

Follow us on:
CVG Group DVL Group

News

18.01.2021

Yaron Lipman (Weizmann Institute of Science) will give a talk in the TUM AI lecture series on Jan 21st, 3pm! Livestream

10.12.2020

Frank Dellaert (Georgia Tech) will give a talk in the TUM AI lecture series on Dec 17th, 4pm! Livestream

15.10.2020

Jon Barron (Google) will give a talk in the TUM AI lecture series on Oct 22nd, 9pm! Livestream

02.10.2020

We have five papers accepted to 3DV 2020!

30.09.2020

Our effcient deep network architectures form the AI engine of the project Slow Down COVID-19 at Harvard.

More


Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision
Previous revision
research:deeplearning [2019/05/22 12:46]
Vladimir Golkov
research:deeplearning [2020/12/01 11:16] (current)
Vladimir Golkov
Line 1: Line 1:
-===== Deep Learning =====+{{linktarget>deeplearning}} 
 +====== Deep Learning ======
  
-**Contact**: [[members:lealtaix|Dr. Laura Leal-Taixe]], [[members:golkov|Vladimir Golkov]], [[members:meinhard|Tim Meinhardt]], [[members:zhouq|Qunjie Zhou]], [[members:dendorfer|Patrick Dendorfer]] 
  
-Deep Learning is a powerful machine learning tool that showed outstanding performance in many fields. One of the greatest successes of Deep Learning has been achieved in large scale object recognition with Convolutional Neural Networks (CNNs). CNNs' main power comes from learning data representations directly from data in a hierarchical layer based structure.+{{:research:deeplearning:ddffnet_cut.png?nolink&700|}}
  
-We apply Convolutional Neural Networks in order to solve computer vision tasks such as optical flow, scene understanding, and develop state-of-the-art methods.+Deep learning is a powerful machine learning framework that has shown outstanding performance in many fields. The main power of deep learning comes from learning data representations directly from data in a hierarchical layer-based structure.
  
-=== Learning by Association === +We apply deep learning to **computer visionautonomous driving, biomedicine, time series data, language, and other fields**, and develop novel methodsAmong the advanced methods we use and develop are **uncertainty quantification, processing of advanced data structures (sequencesgraphs, geometry, high-dimensional data), probabilistic graphical models, reinforcement learning, active learning, domain adaptation, anomaly detection, convolutional networks, recurrent networks, and causality inference.** 
-{{ :research:deeplearning:associative_learning_teaser.png?300 |}} +
-A child is able to learn new concepts quickly and without the need for millions examples that are pointed out individually. Once a child has seen one dogshe or he will be able +
-to recognize other dogs and becomes better at recognition with subsequent exposure to more varietyIn terms of training computers to perform similar tasksdeep neural networks have demonstrated superior performance among machine learning models.+
  
-Howeverthese networks have been trained dramatically differently than a learning child, requiring labels for every training example, following a purely supervised training schemeNeural networks are defined by huge amounts of parameters to be optimized. Thereforea plethora of labeled training data is required, which might be costly and time consuming to obtain. It is desirable to train machine learning models without labels (unsupervisedly) or with only some fraction of the data labeled (semi-supervisedly).+Some of the works from our chair include: [[research:optical_flow_estimation|FlowNet]][[https://hazirbas.com/projects/fusenet/|FuseNet]][[research:vslam:dvso|DVSO]]
  
-We propose a novel training method that follows an intuitive approach: learning by association. We feed a batch of labeled and a batch of unlabeled data through a +==== Contact ====
-network, producing embeddings for both batches. Then, an imaginary walker is sent from samples in the labeled batch to samples in the unlabeled batch. The transition follows a +
-probability distribution obtained from the similarity of the respective embeddings which we refer to as an association.+
  
-In this line of work, we have published papers on **semi-supervised training, domain adaptation, multimodal training (text and images) and unsupervised training clustering**. More information can be found [[https://github.com/haeusser/learning_by_association|here]].+<memberlist> 
 +<dokuwiki> 
 +<filter> 
 +<grps>^deeplearning$</grps> 
 +</filter> 
 +<user>^cremers$</user> 
 +</dokuwiki> 
 +</memberlist>
  
-=== Deep Depth From Focus === +==== Related publications ====
-{{ :research:deeplearning:ddffnet.png?640 |}} +
-DDFF aims at predicting a depth map from a given focal stack in which the focus of the camera gradually changes. DDFFNet is an end-to-end trained Convolutional Neural Network, designed to solve the highly ill-posed //depth from focus// task. +
-Please visit the [[https://hazirbas.com/projects/ddff/|DDFF Project Page]] for details. +
- +
-=== Flownet === +
-In our recent ICCV'15 paper, we presented two CNN architectures to estimate the optical flow given one image pair. We train the network end-to-end on a GPU. Our system works as good as state-of-the-art techniques. +
- +
-<html> <center> +
-<iframe width="640" height="360" src="//www.youtube.com/embed/k_wkDLJ8lJE" frameborder="0" allowfullscreen></iframe> </center> +
-</html> +
- +
-=== Unsupervised Domain Adaptation for Vehicle Control === +
-Even though end-to-end supervised learning has shown promising results for sensorimotor control of self-driving cars, its performance is greatly affected by the weather conditions under which it was trained, showing poor generalization to unseen conditions. Therefore, we show how knowledge can be transferred using semantic maps to new weather conditions without the need to obtain new ground truth data. To this end, we propose to divide the task of vehicle control into two independent modules: a control module which is only trained on one weather condition for which labeled steering data is available, and a perception module which is used as an interface between new weather conditions and the fixed control module. To generate the semantic data needed to train the perception module, we propose to use a generative adversarial network (GAN)-based model to retrieve the semantic information for the new conditions in an unsupervised manner. We introduce a master-servant architecture, where the master model (semantic labels available) trains the servant model (semantic labels not available) +
-===== Related publications =====+
 <bibtex> <bibtex>
-<keywords>deeplearning</keywords>+<keywords>deep learning</keywords>
 <bytype>-1</bytype> <bytype>-1</bytype>
 </bibtex> </bibtex>
 +

Rechte Seite

Informatik IX
Chair of Computer Vision & Artificial Intelligence

Boltzmannstrasse 3
85748 Garching info@vision.in.tum.de

Follow us on:
CVG Group DVL Group

News

18.01.2021

Yaron Lipman (Weizmann Institute of Science) will give a talk in the TUM AI lecture series on Jan 21st, 3pm! Livestream

10.12.2020

Frank Dellaert (Georgia Tech) will give a talk in the TUM AI lecture series on Dec 17th, 4pm! Livestream

15.10.2020

Jon Barron (Google) will give a talk in the TUM AI lecture series on Oct 22nd, 9pm! Livestream

02.10.2020

We have five papers accepted to 3DV 2020!

30.09.2020

Our effcient deep network architectures form the AI engine of the project Slow Down COVID-19 at Harvard.

More