Estimating human pose, shape, and motion from images and video are fundamental challenges with many applications. Recent advances in 2D human pose estimation use large amounts of manually-labeled training data for learning convolutional neural networks (CNNs). Such data is time consuming to acquire and difficult to extend. Moreover, manual labeling of 3D pose, depth and motion is impractical. In this work we present SURREAL: a new large-scale dataset with synthetically-generated but realistic images of people rendered from 3D sequences of human motion capture data. We generate more than 6 million frames together with ground truth pose, depth maps, and segmentation masks. We show that CNNs trained on our synthetic dataset allow for accurate human depth estimation and human part segmentation in real RGB images. Our results and the new dataset open up new possibilities for advancing person analysis using cheap and large-scale synthetic data.
Organizers: Dimitris Tzionas
I will present selected research projects of the Photogrammetry and Remote Sensing Group at ETH, including (i) 3D scene flow estimation for stereo video captured from a car; (ii) extraction of road networks from aerial images; and (iii) 3D reconstruction from large, unstructured (e.g. crowd-sourced) image collections.
The growing scale of image and video datasets in vision makes labeling and annotation of such datasets, for training of recognition models, difficult and time consuming. Further, richer models often require richer labelings of the data, that are typically even more difficult to obtain. In this talk I will focus on two models that make use of different forms of supervision for two different vision tasks.
In the first part of this talk I will focus on object detection. The appearance of an object changes profoundly with pose, camera view and interactions of the object with other objects in the scene. This makes it challenging to learn detectors based on an object-level labels (e.g., “car”). We postulate that having a richer set of labelings (at different levels of granularity) for an object, including finer-grained sub-categories, consistent in appearance and view, and higher-order composites – contextual groupings of objects consistent in their spatial layout and appearance, can significantly alleviate these problems. However, obtaining such a rich set of annotations, including annotation of an exponentially growing set of object groupings, is infeasible. To this end, we propose a weakly-supervised framework for object detection where we discover subcategories and the composites automatically with only traditional object-level category labels as input.
In the second part of the talk I will focus on the framework for large scale image set and video summarization. Starting from the intuition that the characteristics of the two media types are different but complementary, we develop a fast and easily-parallelizable approach for creating not only video summaries but also novel structural summaries of events in the form of the storyline graphs. The storyline graphs can illustrate various events or activities associated with the topic in the form of a branching directed network. The video summarization is achieved by diversity ranking on the similarity graphs between images and video frame, thereby treating consumer image as essentially a form of weak-supervision. The reconstruction of storyline graphs on the other hand is formulated as inference of the sparse time-varying directed graphs from a set of photo streams with assistance of consumer videos.
Time permitting I will also talk about a few other recent project highlights.
Abstract: I will present a general framework for modelling and recovering 3D shape and pose using subdivision surfaces. To demonstrate this frameworks generality, I will show how to recover both a personalized rigged hand model from a sequence of depth images and a blend shape model of dolphin pose from a collection of 2D dolphin images. The core requirement is the formulation of a generative model in which the control vertices of a smooth subdivision surface are parameterized (e.g. with joint angles or blend weights) by a differentiable deformation function. The energy function that falls out of measuring the deviation between the surface and the observed data is also differentiable and can be minimized through standard, albeit tricky, gradient based non-linear optimization from a reasonable initial guess. The latter can often be obtained using machine learning methods when manual intervention is undesirable. Satisfyingly, the "tricks" involved in the former are elegant and widen the applicability of these methods.
In order to avoid an expensive manual labeling process or to learn object classes autonomously without human intervention, object discovery techniques have been proposed that extract visual similar objects from weakly labelled videos. However, the problem of discovering small or medium sized objects is largely unexplored. We observe that videos with activities involving human-object interactions can serve as weakly labelled data for such cases. Since neither object appearance nor motion is distinct enough to discover objects in these videos, we propose a framework that samples from a space of algorithms and their parameters to extract sequences of object proposals. Furthermore, we model similarity of objects based on appearance and functionality, which is derived from human and object motion. We show that functionality is an important cue for discovering objects from activities and demonstrate the generality of the model on three challenging RGB-D and RGB datasets.
In this talk I will discuss two related problems in 3D reconstruction: (i) recovering the 3D shape of a temporally varying non-rigid 3D surface given a single video sequence and (ii) reconstructing different instances of the same object class category given a large collection of images from that category. In both cases we extract dense 3D shape information by analysing shape variation -- in one case of the same object instance over time and in the other across different instances of objects that belong to the same class.
First I will discuss the problem of dense capture of 3D non-rigid surfaces from a monocular video sequence. We take a purely model-free approach where no strong assumptions are made about the object we are looking at or the way it deforms. We apply low rank and spatial smoothness priors to obtain dense non-rigid models using a variational approach.
Second I will describe our recent approach to populating the Pascal VOC dataset with dense, per-object 3D reconstructions, bootstrapped from class labels, ground truth figure-ground segmentations and a small set of keypoint annotations. Our proposed algorithm first estimates camera viewpoint using rigid structure-from-motion, then reconstructs objects shapes by optimizing over visual hull proposals guided by loose within-class shape similarity assumptions.
Even though many challenges remain unsolved, in recent years computer graphics algorithms to render photo-realistic imagery have seen tremendous progress. An important prerequisite for high-quality renderings is the availability of good models of the scenes to be rendered, namely models of shape, motion and appearance. Unfortunately, the technology to create such models has not kept pace with the technology to render the imagery. In fact, we observe a content creation bottleneck, as it often takes man months of tedious manual work by a animation artists to craft models of moving virtual scenes.
To overcome this limitation, the research community has been developing techniques to capture models of dynamic scenes from real world examples, for instance methods that rely on footage recorded with cameras or other sensors. One example are performance capture methods that measure detailed dynamic surface models, for example of actors or an actor's face, from multi-view video and without markers in the scene. Even though such 4D capture methods made big strides ahead, they are still at an early stage of their development. Their application is limited to scenes of moderate complexity in controlled environments, reconstructed detail is limited, and captured content cannot be easily modified, to name only a few restrictions.
In this talk, I will elaborate on some ideas on how to go beyond this limited scope of 4D reconstruction, and show some results from our recent work. For instance, I will show how we can capture more complex scenes with many objects or subjects in close interaction, as well as very challenging scenes of a smaller scale, such a hand motion. The talk will also show how we can capitalize on more sophisticated light transport models and inverse rendering to enable high-quality reconstruction in much more uncontrolled scenes, eventually also outdoors, and with very few cameras. I will also demonstrate how to represent captured scenes such that they can be conveniently modified. If time allows, the talk will cover some of our recent ideas on how to perform advanced edits of videos (e.g. removing or modifying dynamic objects in scenes) by exploiting reconstructed 4D models, as well as robustly found inter- and intra-frame correspondences.
Organizers: Gerard Pons-Moll
A goal in virtual reality is for the user to experience a synthetic environment as if it were real. Engagement with virtual actors is a big part of the sensory context, thus getting the people "right" is critical for success. Size, shape, gender, ethnicity, clothing, color, texture, movement, among other attributes must be layered and nuanced to provide an accurate encounter between an actor and a user. In this talk, I discuss the development of digital human models and how they may be improved to obtain the high realism for successful engagement in a virtual world.
Volumetric 3D modeling has attracted a lot of attention in the past. In this talk I will explain how the standard volumetric formulation can be extended to include semantic information by using a convex multi-label formulation. One of the strengths of our formulation is that it allows us to directly account for the expected surface orientations. I will focus on two applications. Firstly, I will introduce a method that allows for joint volumetric reconstruction and class segmentation. This is achieved by taking into account the expected orientations of object classes such as ground and building. Such a joint approach considerably improves the quality of the geometry while at the same time it gives a consistent semantic segmentation. In the second application I will present a method that allows for the reconstruction of challenging objects such as for example glass bottles. The main difficulty with reconstructing such objects are the texture-less, transparent and reflective areas in the input images. We propose to formulate a shape prior based on the locally expected surface orientation to account for the ambiguous input data. Our multi-label approach also directly enables us to segment the object from its surrounding.
The goal of lifelong visual learning is to develop techniques that continuously and autonomously learn from visual data, potentially for years or decades. During this time the system should build an ever-improving base of generic visual information, and use it as background knowledge and context for solving specific computer vision tasks. In my talk, I will highlight two recent results from our group on the road towards lifelong visual scene understanding: the derivation of theoretical guarantees for lifelong learning systems and the development of practical methods for object categorization based on semantic attributes.
Organizers: Gerard Pons-Moll
Point-light walkers and stick figures rendered orthographically and without self-occlusion do not contain any information as to their depth. For instance, a frontoparallel projection could depict a walker from the front or from the back. Nevertheless, observers show a strong bias towards seeing the walker as facing the viewer. A related stimulus, the silhouette of a human figure, does not seem to show such a bias. We develop these observations into a tool to study the cause of the facing the viewer bias observed for biological motion displays.
I will give a short overview about existing theories with respect to the facing-the-viewer bias, and about a number of findings that seem hard to explain with any single one of them. I will then present the results of our studies on both stick figures and silhouettes which gave rise to a new theory about the facing the viewer bias, and I will eventually present an experiment that tests a hypothesis resulting from it. The studies are discussed in the context of one of the most general problems the visual system has to solve: How do we disambiguate an initially ambiguous sensory world and eventually arrive at the perception of a stable, predictable "reality"?