This year’s ECCV 2018 conference experienced an unprecedented growth of community and brought to light the most recent advances in computer vision. As expected, all the sessions were dominated by Deep Learning with Convolutional Neural Networks (CNNs).

For those who couldn’t join, I picked up a few interesting topics that caught my attention. Here is the list:

Autonomous Driving

Self-localization on-the-fly

One of the main topics at ECCV 2018 was autonomous driving. Can you compete against LIDAR? Can you detect and reconstruct cars as 3D objects from video? Check some ECCV’s challenges!

 

CARLA: Democratizing Autonomous Driving Research

Autonomous driving requires extremely safe behavior with respect to urban areas. Have you heard of CARLA Simulator? The tutorial can be found here.

Novel Neural Network Architectures and Schemes

Convolutional Networks with Adaptive Computation Graphs

Anreas Veit is showing that CNNs don’t necessarily need fixed feed-forward structure. Instead, he is proposing adaptive network topology which outperforms ResNet on ImageNet. You may see paper summary here.

Lifting Layers: Analysis and Application

How can increasing the dimensionality of an input help in deep learning applications in the fields of image classification or denoising?  What about faster training?

Jointly Discovering Visual Objects and Spoken Words from Raw Sensory Input

Did you know neural networks are able to discover audio-visual semantic correspondence so that we may highlight objects, scenes or regions (in the image) that someone is talking about?

Learning Discriminative Video Representations Using Adversarial Perturbations

CNN features, Stiefel manifold, Riemannian conjugate gradient scheme – all the geeky stuff at once.

Magic with Images

Diverse Image-to-Image Translation via Disentangled Representations

How can we generate a new images with specific content and attribute?

Style-aware Content Loss for Real-time HD Style Transfer

Authors lay the groundwork for the problem of style transfer for images. They provide us with deceptive “paintings” which mimic various historical artists’ styles. These are generated by CNNs from real world images and achieve high deception rate.  This works even for video!

Kochen mit Spaß!

I found out there is the Cookpad mobile app! You may use it to take a picture of a new delicious meal and CNN will identify the meal so that the app can tell you what ingredients you have to get and how to cook the meal on your own.

Computer Vision community is growing rapidly as well as all the number of brilliant ideas and applications!