Raimar Wagner

Learning convolutional neural networks from few samples

Learning Convolutional Neural Networks (CNN) is commonly carried out by plain supervised gradient descent. With sufficient training data, this leads to very competitive results for visual recognition tasks when starting from a random initialization. When the amount of labeled data is limited, CNNs reveal their strong dependence on large amounts of training data. However, recent results have shown that a well chosen optimization starting point can be beneficial for convergence to a good generalizing minimum. This starting point was mostly found using unsupervised feature learning techniques such as sparse coding or transfer learning from related recognition tasks. In this work, we compare these two approaches against a simple patch based initialization scheme and a random initialization of the weights. We show that pre-training helps to train CNNs from few samples and that the correct choice of the initialization scheme can push the networks performance by up to 41% compared to random initialization.

High-performance on-road vehicle detection in monocular images

This paper addresses the problem of monocular vehicle detection for forward collision warning. We present a system that is able to process large images with high speed and delivers high detection rates at only one false alarm every 100 frames.

Feature selection for automotive object detection tasks - A study

For object detection in monocular images, the Boosted Cascade [1] has become the standard approach for driver assistance systems. This paper studies the discriminative power of different features for common automotive object detection tasks: pedestrian and vehicle detection using infrared cameras at night, as well as pedestrian and vehicle detection in daylight conditions. It is shown that the use of intra-stage information propagation with Activation History Features (AHFs) [2] significantly speeds up the detection at the same detection rate. Thus, AHFs offer a speedup at no cost.

Improving detector performance by learning from compressed samples

Increasing data volumes coupled with bandwidth limitations in on-board data transmission paths make data compression of automotive video signals indispensable. Since traditional image compression algorithms are solely tuned for optimal human perception, this work studies their effect on a nighttime automotive pedestrian detection system. Evaluating raw-data trained detectors on compressed video streams reveals detection rate declines for strong compression factors. On the other hand, when using image compression as training data preprocessing tool an increase in detection performance can be achieved.

Model-based video compression for real world data

This work examines the benefits for displaying video content on a 2D screen that can be gained by incorporating depth information into the encoding process. The creation of a 3D model helps to compensate for redundant encoding of reappearing contents beyond group of picture borders, like needed in hybrid video codecs. Starting with filtering the image data, creating a 3D model and applying a model-based compression via MPEG-IV / Part 25 to the model, finally a 2D video representation is rendered. Subjective quality assessments of first preliminary results show, that the perceived video quality is comparable with H.264/AVC encoded videos for low quality scenarios when working with a static 3D model.

Night time road curvature estimation based on Convolutional Neural Networks

Detecting the road geometry at night time is an essential precondition to provide optimal illumination for the driver and the other traffic participants. In this paper we propose a novel approach to estimate the current road curvature based on three sensors: A far infrared camera, a near infrared camera and an imaging radar sensor. Various Convolutional Neural Networks with different configuration are trained for each input. By fusing the classifier responses of all three sensors, a further performance gain is achieved. To annotate the training and evaluation dataset without costly human interaction a fully automatic curvature annotation algorithm based on inertial navigation system is presented as well.

Convolutional Neural Networks for night-time animal orientation estimation

In rural areas, wildlife animal road crossings are a threat to both the driver and the wildlife population. Since most accidents take place at night, recent night vision driver assistance systems are supporting the driver by automatically detecting animals on infrared camera imagery. After detecting an animal on the roadside, the orientation towards the road can give a first cue for an upcoming trajectory prediction. This paper describes an novel classification-based scheme for nighttime animal orientation estimation from single infrared images. Our system classifies already detected animals, in particular deer, as being either oriented left, right or back/front. We propose an approach based on Convolutional Neural Networks which learns multiple stages of invariant features in a supervised end-to-end fashion. Experiments show that our method outperforms baseline methods like HOG/SVM or boosted Haar-stumps on this task.

Video and disparity compression for automotive stereo vision algorithms

This paper introduces a compression system for automotive stereo algorithms. We propose a method to compress video, disparity and occlusion information based on a 12 bit per pixel H.264/AVC encoder. A rate allocation parameter controls the available bit rate between video and disparity. In addition, the occlusion information is lossless encoded into the bit stream. For driver assistance systems that rely on the disparity information, a higher bit rate can be allocated for the disparity to obtain a better accuracy.

Model optimization for model-based compression of real world video data

The availability of affordable cameras with additional depth sensor boosted the advances in the field of Simultaneous Localization and Mapping (SLAM). This area of research addresses creating a 3D representation of unknown environment by fusing sensor information. Whereas a coarse representation suffices for navigational purposes, the resulting 3D models are not accurate enough to allow rendering a visually pleasing video stream based on the gathered information, when applying a model based video compression standard such as MPEG-4 Part 25. This publication focuses on increasing the accuracy of the created 3D models in order to improve the perceived quality when using model-based video coding and aims for performance comparable to hybrid video encoding.

Image compression with discriminative dictionaries

Common image compression algorithms like JPEG or JPEG2000 transform the individual pixel values into a domain that favors a compact representation. In contrast to the fixed DCT or Wavelet domains, recent efforts were made on image coding with learned overcomplete dictionaries. In this work, we investigate the question whether dictionaries based on classification features are usable for image compression. We show that, despite their original purpose is to extract discriminative features within a Convolutional Neural Network, these features are capable of reaching competitive compression results when combined with a sparsity promoting coding scheme.