Mgr. Ondřej Drbohlav, Ph.D.

The Visual Object Tracking challenge VOT2021 is the ninth annual tracker benchmarking activity organized by the VOT initiative. Results of 71 trackers are presented; many are state-of-the-art trackers published at major computer vision conferences or in journals in recent years. The VOT2021 challenge was composed of four sub-challenges focusing on different tracking domains: (i) VOT-ST2021 challenge focused on short-term tracking in RGB, (ii) VOT-RT2021 challenge focused on "real-time" short-term tracking in RGB, (iii) VOT-LT2021 focused on long-term tracking, namely coping with target disappearance and reappearance and (iv) VOT-RGBD2021 challenge focused on long-term tracking in RGB and depth imagery. The VOT-ST2021 dataset was refreshed, while VOT-RGBD2021 introduces a training dataset and sequestered dataset for winner identification. The source code for most of the trackers, the datasets, the evaluation kit and the results along with the source code for most trackers are publicly available at the challenge website.

The Visual Object Tracking challenge VOT2020 is the eighth annual tracker benchmarking activity organized by the VOT initiative. Results of 58 trackers are presented; many are state-of-the-art trackers published at major computer vision conferences or in journals in the recent years. The VOT2020 challenge was composed of five sub-challenges focusing on different tracking domains: (i) VOT-ST2020 challenge focused on short-term tracking in RGB, (ii) VOT-RT2020 challenge focused on “real-time” short-term tracking in RGB, (iii) VOT-LT2020 focused on long-term tracking namely coping with target disappearance and reappearance, (iv) VOT-RGBT2020 challenge focused on short-term tracking in RGB and thermal imagery and (v) VOT-RGBD2020 challenge focused on long-term tracking in RGB and depth imagery. Only the VOT-ST2020 datasets were refreshed. A significant novelty is introduction of a new VOT short-term tracking evaluation methodology, and introduction of segmentation ground truth in the VOT-ST2020 challenge – bounding boxes will no longer be used in the VOT-ST challenges. A new VOT Python toolkit that implements all these novelites was introduced. Performance of the tested trackers typically by far exceeds standard baselines. The source code for most of the trackers is publicly available from the VOT page. The dataset, the evaluation kit and the results are publicly available at the challenge website (http://votchallenge.net).

The Visual Object Tracking challenge VOT2019 is the seventh annual tracker benchmarking activity organized by the VOT initiative. Results of 81 trackers are presented; many are state-of-the-art trackers published at major computer vision conferences or in journals in the recent years. The evaluation included the standard VOT and other popular methodologies for short-term tracking analysis as well as the standard VOT methodology for long-term tracking analysis. The VOT2019 challenge was composed of five challenges focusing on different tracking domains: (i) VOTST2019 challenge focused on short-term tracking in RGB, (ii) VOT-RT2019 challenge focused on "real-time" shortterm tracking in RGB, (iii) VOT-LT2019 focused on longterm tracking namely coping with target disappearance and reappearance. Two new challenges have been introduced: (iv) VOT-RGBT2019 challenge focused on short-term tracking in RGB and thermal imagery and (v) VOT-RGBD2019 challenge focused on long-term tracking in RGB and depth imagery. The VOT-ST2019, VOT-RT2019 and VOT-LT2019 datasets were refreshed while new datasets were introduced for VOT-RGBT2019 and VOT-RGBD2019. The VOT toolkit has been updated to support both standard shortterm, long-term tracking and tracking with multi-channel imagery. Performance of the tested trackers typically by far exceeds standard baselines. The source code for most of the trackers is publicly available from the VOT page. The dataset, the evaluation kit and the results are publicly available at the challenge website.

We revisit the problem of local optimization (LO) in RANSAC for homography estimation. The standard state-of-the-art LO-RANSAC improves the plain version's accuracy and stability, but it may be computationally demanding, it is complex to implement and requires setting multiple parameters. We show that employing L1 minimization instead of the standard LO step of LO-RANSAC leads to results with similar precision. At the same time, the proposed L1 minimization is significantly faster than the standard LO step of [8], it is easy to implement and it has only a few of parameters which all have intuitive interpretation. On the negative side, the L1 minimization does not achieve the robustness of the standard LO step, its probability of failure is higher.

A conceptually very simple unsupervised algorithm for learning structure in the form of a hierarchical probabilistic model is described in this paper. The proposed probabilistic model can easily work with any type of image primitives such as edge segments, non-max-suppressed filter set responses, texels, distinct image regions, SIFT features, etc., and is even capable of modelling non-rigid and/or visually variable objects. The model has recursive form and consists of sets of simple and gradually growing sub-models that are shared and learned individually in layers. The proposed probabilistic framework enables to exactly compute the probability of presence of a certain model, regardless on which layer it actually is. All these learned models constitute a rich set of independent structure elements of variable complexity that can be used as features in various recognition tasks.

The long term goal of artificial intelligence and computer vision is to be able to build models of the world automatically and to use them for interpretation of new situations. It is natural that such models are efficiently organized in a hierarchical manner; a model is build by sub-models, these sub-models are again build of another models, and so on. These building blocks are usually shareable; different objects may consist of the same components. In this paper, we describe a hierarchical probabilistic model for visual domain and propose a method for its efficient inference based on data partitioning and dynamic programming. We show the behaviour of the model, which is in this case made manually, and inference method on a controlled yet challenging dataset consisting of rotated, scaled and occluded letters. The experiments show that the proposed model is robust to all above-mentioned aspects.

The most usual ways of observation of satellites and space debris and measur ement of their orbits are telescopic images, radar reflections, laser measurements. We use two of these three modalities, we combine telescopic images with response of Doppler radars. We u se single images from a terrestrial telescope. Our radar is passive, we receive the signal of a distant terrestrial transmitter. The receiver has a non-directional antenna and only Doppler shift is employed to gain information about an object's movement. Due to sensitivity limitati ons, our approach is applicable to large objects (RCS 5 m2 ) at distances 2000 km. Our method requires simultaneous detections by a telescope and a radar during the same fly-over, not necessarily at exactly the same time.

Our goal is to identify hierarchical compositional models from highly cluttered data. The data to learn from are assumed to be imperfect in two respects. Firstly, large portion of the data is coming from background clutter. Secondly, data generated by a recursive compositional model are subject to random replacements of correct descendants by randomly chosen ones at every level of the hierarchy. In this paper, we study the limits and capabilities of an approach which is based on likelihood maximization. The algorithm makes explicit probabilistic assignments of individual data to compositional model and background clutter. It uses these assignments to effectively focus on the data coming from the compositional model and iteratively estimate their compositional structure.

In this paper, we test the performance of a hierarchical shape detector on the problem of window detection in facade images. The hierarchical shape model detector is constructed automatically using a small number of hand-drawn images of windows. The window detections are evaluated on both rectified and non-rectified facade images. On an eTRIMS dataset containing around 1000 windows, the detector found around 600 windows and 250 false detections in rectified images. Similar performance was obtained for non-rectified facades.

3D texture classification under varying viewpoint and illumination has been a vivid research topic, and many methods have been developed. It is crucial that these methods be compared using an unbiased evaluation methodology. The most frequently employed methodologies use images from the Columbia-Utrecht Reflectance and Texture Database. These methodologies construct the training and test sets to be disjoint in the imaging parameters, but do not separate them spatially because they use images of the same surface patch for both. We perform a series of experiments which show that such practice leads to overestimation of classifier performance and distorts experimental findings. To correct that, we accurately register the images across all imaging conditions and split the surface patches to parts. The training and testing is then done on spatially disjoint parts. We show that such methodology gives a more realistic assessment of classifier performance. The sample annotations for all images

Helmholtz stereopsis is a relatively recent reconstruction technique which is able to reconstruct scenes with arbitrary and unknown surface reflectance properties. Conventional implementations of the method estimate surface normal direction at each surface point via an eigenanalysis, thereby optimising an algebraic distance. We develop a more physically meaningful radiometric distance whose minimisation is shown to yield a Maximum Likelihood surface normal estimate. The proposed method produces more accurate results than algebraic methods on synthetic imagery and yields excellent reconstruction results on real data. Our analysis explains why, for some imaging configurations, a sub-optimal algebraic distance can yield good results.

We report results from a new methodology for investigating the visually perceived properties of surface textures. Densely sampled two-dimensional 1/f^beta-noise processes are used to model natural looking surfaces, which are rendered using combined point-source and ambient lighting. Surfaces are shown in motion to provide rich cues to their relief. They are generated in real time to enable observers to dynamically manipulate surface parameters. A method of adjustment is employed to investigate the effects that the two surface parameters, magnitude roll-off factor and RMS height, have on perceived roughness. The results are used to develop an estimation method for perceived roughness.

So far, Helmholtz Stereopsis has been widely applied to objects of smooth geometry and piecewise uniform Bidirectional Reflectance Distribution Function (BRDF). Moreover, for non-convex surfaces the inter-reflection effects have been completely neglected. This paper extends the method to surfaces which exhibit strong texture, nontrivial geometry, and are possibly non-convex.

Radiometric calibration of a Helmholtz stereo setup is presented that does not require external light calibration. The problem is shown ill-posed for the case of two cameras but well posed for a greater number of cameras. A calibration procedure that includes a regularizer for the two-camera case is proposed.

Specularities Reduce Ambiguity of Uncalibrated Photometric Stereo