Deep learning has significantly advanced state-of-the-art for 3D hand pose estimation, of which accuracy can be improved with increased amounts of labelled data. However, acquiring 3D hand pose labels can be extremely difficult. In this talk, I will present our recent two works on leveraging self-supervised learning techniques for hand pose estimation from depth map. In both works, we incorporate differentiable renderer to the network and formulate training loss as model fitting error to update network parameters. In first part of the talk, I will present our earlier work which approximates hand surface with a set of spheres. We then model the pose prior as a variational lower bound with variational auto-encoder(VAE). In second part, I will present our latest work on regressing the vertex coordinates of a hand mesh model with 2D fully convolutional network(FCN) in a single forward pass. In the first stage, the network estimates a dense correspondence field for every pixel on the image grid to the mesh grid. In the second stage, we design a differentiable operator to map features learned from the previous stage and regress a 3D coordinate map on the mesh grid. Finally, we sample from the mesh grid to recover the mesh vertices, and fit it an articulated template mesh in closed form. Without any human annotation, both works can perform competitively with strongly supervised methods. The later work will also be later extended to be compatible with MANO model.
Organizers: Dimitrios Tzionas
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"?
Compared to static image segmentation, video segmentation is still in its infancy. Various research groups have different tasks in mind when they talk of video segmentation. For some it is motion segmentation, some think of an over-segmentation with thousands of regions per video, and others understand video segmentation as contour tracking. I will go through what I think are reasonable video segmentation subtasks and will touch the issue of benchmarking. I will also discuss the difference between image and video segmentation. Due to the availability of motion and the redundancy of successive frames, video segmentation should actually be easier than image segmentation. However, recent evidence indicates the opposite: at least at the level of superpixel segmentation, image segmentation methodology is more advanced than what can be found in the video segmentation literature.
Organizers: Gerard Pons-Moll
In the first part of our talk, we present an approach for large displacement optical flow. Optical flow computation is a key component in many computer vision systems designed for tasks such as action
detection or activity recognition. Inspired by the large displacement optical flow of Brox and Malik, our approach DeepFlow combines a novel matching algorithm with a variational approach . Our matching algorithm builds upon a multi-stage architecture interleaving convolutions and max-pooling. DeepFlow efficiently handles large displacements occurring in realistic videos, and shows competitive performance on optical flow benchmarks.
In the second part of our talk, we present a state-of-the-art approach for action recognition based on motion stabilized trajectory descriptors and a Fisher vector representation. We briefly review the recent trajectory-based video features and, then, introduce their motion stabilized version, combining human detection and dominant motion estimation. Fisher vectors summarize the information of a video efficiently. Results on several of the recent action datasets as well as the TrecVid MED dataset show that our approach outperforms the state-of-the-art
Computer vision problems often involve optimization of two quantities, one of which is time. Such problems can be formulated as time-constrained optimization or performance-constrained search for the fastest algorithm. We show that it is possible to obtain quasi-optimal time-constrained solutions to some vision problems by applying Wald's theory of sequential decision-making. Wald assumes independence of observation, which is rarely true in computer vision. We address the problem by combining Wald's sequential probability ratio test and AdaBoost. The solution, called the WaldBoost, can be viewed as a principled way to build a close-to-optimal “cascade of classifiers” of the Viola-Jones type. The approach will be demonstrated on four tasks: (i) face detection, (ii) establishing reliable correspondences between image, (iii) real-time detection of interest points and (iv) model search and outlier detection using RANSAC. In the face detection problem, the objective is learning the fastest detector satisfying constraints on false positive and false negative rates. The correspondence pruning addresses the problem of fast selection with a predefined false negative rated. In interest point problem we show how a fast implementation of known detectors can obtained by Waldboost. The “mimicked” detectors provide a training set of positive and negative examples of interest ponts and WaldBoost learns a detector, (significantly) faster than the providers of the training set, formed as a linear combination of efficiently computable feature. In RANSAC, we show how to exploit Wald's test in a randomised model verification procedure to obtain an algorithm significantly faster than deterministic verification yet with equivalent probabilistic guarantees of correctness.
Organizers: Gerard Pons-Moll
Stereo matching -- establishing correspondences between images taken from nearby viewpoints -- is one of the oldest problems in computer vision. While impressive progress has been made over the last two decades, most current stereo methods do not scale to the high-resolution images taken by today's cameras since they require searching the full space of all possible disparity hypotheses over all pixels.
In this talk I will describe a new scalable stereo method that only evaluates a small portion of the search space. The method first generates plane hypotheses from matched sparse features, which are then refined into surface hypotheses using local slanted plane sweeps over a narrow disparity range. Finally, each pixel is assigned to one of the local surface hypotheses. The technique achieves significant speedups over previous algorithms and achieves state-of-the-art accuracy on high-resolution stereo pairs of up to 19 megapixels.
I will also present a new dataset of high-resolution stereo pairs with subpixel-accurate ground truth, and provide a brief outlook on the upcoming new version of the Middlebury stereo benchmark.
This talk will give an overview of some of the research in the Image and Video Computing Group at Boston University related to image- and video-based analysis of humans and their behavior, including: tracking humans, localizing and classifying actions in space-time, exploiting contextual cues in action classification, estimating human pose from images, analyzing the communicative behavior of children in video, and sign language recognition and retrieval.
Collaborators in this work include (in alphabetical order): Vassilis Athitsos, Qinxun Bai, Margrit Betke, R. Gokberk Cinbis, Kun He, Nazli Ikizler-Cinbis, Hao Jiang, Liliana Lo Presti, Shugao Ma, Joan Nash, Carol Neidle, Agata Rozga, Tai-peng Tian, Ashwin Thangali, Zheng Wu, and Jianming Zhang.