Video Object Segmentation, and video processing in general, has been historically dominated by methods that rely on the temporal consistency and redundancy in consecutive video frames. When the temporal smoothness is suddenly broken, such as when an object is occluded, or some frames are missing in a sequence, the result of these methods can deteriorate significantly or they may not even produce any result at all. This paper explores the orthogonal approach of processing each frame independently, i.e disregarding the temporal information. In particular, it tackles the task of semi-supervised video object segmentation: the separation of an object from the background in a video, given its mask in the first frame. We present Semantic One-Shot Video Object Segmentation (OSVOS-S), based on a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation, and finally to learning the appearance of a single annotated object of the test sequence (hence one shot). We show that instance level semantic information, when combined effectively, can dramatically improve the results of our previous method, OSVOS. We perform experiments on two recent video segmentation databases, which show that OSVOS-S is both the fastest and most accurate method in the state of the art.
This paper introduces a novel algorithm for transductive inference in higher-order MRFs, where the unary energies are parameterized by a variable classifier. The considered task is posed as a joint optimization problem in the continuous classifier parameters and the discrete label variables. In contrast to prior approaches such as convex relaxations, we propose an advantageous decoupling of the objective function into discrete and continuous subproblems and a novel, efficient optimization method related to ADMM. This approach preserves integrality of the discrete label variables and guarantees global convergence to a critical point. We demonstrate the advantages of our approach in several experiments including video object segmentation on the DAVIS data set and interactive image segmentation.
In order to track all persons in a scene, the tracking-by-detection paradigm has proven to be a very effective approach. Yet, relying solely on a single detector is also a major limitation, as useful image information might be ignored. Consequently, this work demonstrates how to fuse two detectors into a tracking system. To obtain the trajectories, we propose to formulate tracking as a weighted graph labeling problem, resulting in a binary quadratic program. As such problems are NP-hard, the solution can only be approximated. Based on the Frank-Wolfe algorithm, we present a new solver that is crucial to handle such difficult problems. Evaluation on pedestrian tracking is provided for multiple scenarios, showing superior results over single detector tracking and standard QP-solvers. Finally, our tracker ranks 2nd on the MOT16 benchmark and 1st on the new MOT17 benchmark, outperforming over 90 trackers.
We present VI-DSO, a novel approach for visual-inertial odometry, which jointly estimates camera poses and sparse scene geometry by minimizing photometric and IMU measurement errors in a combined energy functional. The visual part of the system performs a bundle-adjustment like optimization on a sparse set of points, but unlike key-point based systems it directly minimizes a photometric error. This makes it possible for the system to track not only corners, but any pixels with large enough intensity gradients. IMU information is accumulated between several frames using measurement preintegration, and is inserted into the optimization as an additional constraint between keyframes. We explicitly include scale and gravity direction into our model and jointly optimize them together with other variables such as poses. As the scale is often not immediately observable using IMU data this allows us to initialize our visual-inertial system with an arbitrary scale instead of having to delay the initialization until everything is observable. We perform partial marginalization of old variables so that updates can be computed in a reasonable time. In order to keep the system consistent we propose a novel strategy which we call "dynamic marginalization". This technique allows us to use partial marginalization even in cases where the initial scale estimate is far from the optimum. We evaluate our method on the challenging EuRoC dataset, showing that VI-DSO outperforms the state of the art.
The finding that very large networks can be trained efficiently and reliably has led to a paradigm shift in computer vision from engineered solutions to learning formulations. As a result, the research challenge shifts from devising algorithms to creating suitable and abundant training data for supervised learning. How to efficiently create such training data? The dominant data acquisition method in visual recognition is based on web data and manual annotation. Yet, for many computer vision problems, such as stereo or optical flow estimation, this approach is not feasible because humans cannot manually enter a pixel-accurate flow field. In this paper, we promote the use of synthetically generated data for the purpose of training deep networks on such tasks.We suggest multiple ways to generate such data and evaluate the influence of dataset properties on the performance and generalization properties of the resulting networks. We also demonstrate the benefit of learning schedules that use different types of data at selected stages of the training process.
Micro aerial vehicles (MAVs) are strongly limited in their payload and power capacity. In order to implement autonomous navigation, algorithms are therefore desirable that use sensory equipment that is as small, low-weight, and low-power consuming as possible. In this paper, we propose a method for autonomous MAV navigation and exploration using a low-cost consumer-grade quadrocopter equipped with a monocular camera. Our vision-based navigation system builds on LSD-SLAM which estimates the MAV trajectory and a semi-dense reconstruction of the environment in real-time. Since LSD-SLAM only determines depth at high gradient pixels, texture-less areas are not directly observed so that previous exploration methods that assume dense map information cannot directly be applied. We propose an obstacle mapping and exploration approach that takes the properties of our semi-dense monocular SLAM system into account. In experiments, we demonstrate our vision-based autonomous navigation and exploration system with a Parrot Bebop MAV.
We present a novel preconditioning technique for proximal optimization methods that relies on graph algorithms to construct effective preconditioners. Such combinatorial preconditioners arise from partitioning the graph into forests. We prove that certain decompositions lead to a theoretically optimal condition number. We also show how ideal decompositions can be realized using matroid partitioning and propose efficient greedy variants thereof for large-scale problems. Coupled with specialized solvers for the resulting scaled proximal subproblems, the preconditioned algorithm achieves competitive performance in machine learning and vision applications.
We propose proximal backpropagation (ProxProp) as a novel algorithm that takes implicit instead of explicit gradient steps to update the network parameters during neural network training. Our algorithm is motivated by the step size limitation of explicit gradient descent, which poses an impediment for optimization. ProxProp is developed from a general point of view on the backpropagation algorithm, currently the most common technique to train neural networks via stochastic gradient descent and variants thereof. Specifically, we show that backpropagation of a prediction error is equivalent to sequential gradient descent steps on a quadratic penalty energy, which comprises the network activations as variables of the optimization. We further analyze theoretical properties of ProxProp and in particular prove that the algorithm yields a descent direction in parameter space and can therefore be combined with a wide variety of convergent algorithms. Finally, we devise an efficient numerical implementation that integrates well with popular deep learning frameworks. We conclude by demonstrating promising numerical results and show that ProxProp can be effectively combined with common first order optimizers such as Adam.
We tackle the problem of reflectance estimation from a set of multi-view images, assuming known geometry. The approach we put forward turns the input images into reflectance maps, through a robust variational method. The variational model comprises an image-driven fidelity term and a term which enforces consistency of the reflectance estimates with respect to each view. If illumination is fixed across the views, then reflectance estimation remains under-constrained: a regularization term, which ensures piecewise-smoothness of the reflectance, is thus used. Reflectance is parameterized in the image domain, rather than on the surface, which makes the numerical solution much easier, by resorting to an alternating majorization-minimization approach. Experiments on both synthetic and real datasets are carried out to validate the proposed strategy.
Visual scene understanding is an important capability that enables robots to purposefully act in their environment. In this paper, we propose a novel approach to object-class segmentation from multiple RGB-D views using deep learning. We train a deep neural network to predict object-class semantics that is consistent from several view points in a semi-supervised way. At test time, the semantics predictions of our network can be fused more consistently in semantic keyframe maps than predictions of a network trained on individual views. We base our network architecture on a recent single-view deep learning approach to RGB and depth fusion for semantic object-class segmentation and enhance it with multi-scale loss minimization. We obtain the camera trajectory using RGB-D SLAM and warp the predictions of RGB-D images into ground-truth annotated frames in order to enforce multi-view consistency during training. At test time, predictions from multiple views are fused into keyframes. We propose and analyze several methods for enforcing multi-view consistency during training and testing. We evaluate the benefit of multi-view consistency training and demonstrate that pooling of deep features and fusion over multiple views outperforms single-view baselines on the NYUDv2 benchmark for semantic segmentation. Our end-to-end trained network achieves state-of-the-art performance on the NYUDv2 dataset in single-view segmentation as well as multi-view semantic fusion.