Dealing properly with different viewing conditions remains a key challenge for computer vision in autonomous driving. Domain adaptation has opened new possibilities for data augmentation, translating arbitrary road scene images into different environmental conditions. Although multimodal concepts have demonstrated the capability to separate content and style, we find that existing methods fail to reproduce scenes in the exact appearance given by a reference image. In this paper, we address the aforementioned problem by introducing a style alignment loss between output and reference image. We integrate this concept into a multimodal unsupervised image-to-image translation model with a novel dual-mode training process and additional adversarial losses. Focusing on road scene images, we evaluate our model in various aspects including visual quality and feature matching. Our experiments reveal that we are able to significantly improve both style alignment and image quality in different viewing conditions. Adapting concepts from neural style transfer, our new training approach allows to control the output of multimodal domain adaptation, making it possible to generate arbitrary scenes and viewing conditions for data augmentation.

We present a novel image classification technique for detecting multiple objects (called subobjects) in a single image. In addition to image classifiers, we apply spatial relationships among the subobjects to verify and to predict the locations of detected and undetected subobjects, respectively. By continuously refining the spatial relationships throughout the detection process, even locations of completely occluded exhibits can be determined. This approach is applied in the context of PhoneGuide, an adaptive museum guidance system for camera-equipped mobile phones.

In this paper, we present a novel technique for adapting local image classifiers that are applied for object recognition on mobile phones through ad-hoc network communication between the devices. By continuously accumulating and exchanging collected user feedback among mobile phones that are located within signal range, we show that our approach improves the overall classification rate and adapts to dynamic changes quickly. This technique is applied in the context of PhoneGuide -- an adaptive museum guidance system.

In this paper, we present a high speed optical motion capture method that can measure three dimensional motion, orientation, and incident illumination at tagged points in a scene. We use tracking tags that work in natural lighting conditions and can be imperceptibly embedded in attire or other objects. Our system supports an unlimited number of tags in a scene, with each tag uniquely identified to eliminate marker reacquisition issues. Our tags also provide incident illumination data which can be used to match scene lighting when inserting synthetic elements. The technique is therefore ideal for on-set motion capture or real-time broadcasting of virtual sets. Unlike previous methods that employ high speed cameras or scanning lasers, we capture the scene appearance using the simplest possible optical devices - a light-emitting diode (LED) with a passive binary mask used as the transmitter and a photosensor used as the receiver. We strategically place a set of optical transmitters to spatio-temporally encode the volume of interest. Photosensors attached to scene points demultiplex the coded optical signals from multiple transmitters, allowing us to compute not only receiver location and orientation but also their incident illumination and the reflectance of the surfaces to which the photosensors are attached. We use our untethered tag system, called Prakash, to demonstrate methods of adding special effects to captured videos that cannot be accomplished using pure vision techniques that rely on camera images.

We present PhoneGuide -- an enhanced museum guidance system that uses camera-equipped mobile phones and on-device object recognition.Our main technical achievement is a simple and light-weight object recognition approach that is realized with single-layer perceptron neuronal networks. In contrast to related systems which perform computationally intensive image processing tasks on remote servers, our intention is to carry out all computations directly on the phone. This ensures little or even no network traffic and consequently decreases cost for online times. Our laboratory experiments and field surveys have shown that photographed museum exhibits can be recognized with a probability of over 90%.We have evaluated different feature sets to optimize the recognition rate and performance. Our experiments revealed that normalized color features are most effective for our method. Choosing such a feature set allows recognizing an object below one second on up-to-date phones. The amount of data that is required for differentiating 50 objects from multiple perspectives is less than 6KBytes.