We propose VLOGGER, a method for audio-driven human video generation from a single input image of a person, which builds on the success of recent generative diffusion models. Our method consists of 1) a stochastic human-to-3d-motion diffusion model, and 2) a novel diffusion-based architecture that augments text-to-image models with both spatial and temporal controls. This supports the generation of high quality video of variable length, easily controllable through text or speech via high-level representations of human faces and bodies. In contrast to previous work, our method does not require training for each person, does not rely on face detection and cropping, generates the complete image (not just the face or the lips), and considers a broad spectrum of scenarios (e.g. visible torso or diverse subject identities) that are critical to correctly synthesize humans who communicate. We also curate MENTOR, a new and diverse dataset with 3d pose and expression annotations, one order of magnitude larger than previous ones (800,000 identities) and with dynamic gestures, where we train and ablate our main technical contributions. VLOGGER outperforms state-of-the-art methods in three public benchmarks, considering image quality, identity preservation and temporal consistency while also generating upperbody gestures. We analyze the performance of VLOGGER with respect to multiple diversity metrics, showing that our architectural choices and the use of MENTOR benefit training a fair and unbiased model at scale. Finally we show applications in video editing and personalization.

Understanding 3D human interactions is fundamental for fine-grained scene analysis and behavioural modeling. However, most of the existing models predict incorrect, lifeless 3D estimates, that miss the subtle human contact aspects-the essence of the event-and are of little use for detailed behavioral understanding. This paper addresses such issues with several contributions: (1) we introduce models for interaction signature estimation (ISP) encompassing contact detection, segmentation, and 3D contact signature prediction; (2) we show how such components can be leveraged to ensure contact consistency during 3D reconstruction; (3) we construct several large datasets for learning and evaluating 3D contact prediction and reconstruction methods; specifically, we introduce CHI3D, a lab-based accurate 3D motion capture dataset with 631 sequences containing 2,525 contact events, 728,664 ground truth 3D poses, as well as FlickrCI3D, a dataset of 11,216 images, with 14,081 processed pairs of people, and 81,233 facet-level surface correspondences. Finally, (4) we propose methodology for recovering the ground-truth pose and shape of interacting people in a controlled setup and (5) annotate all 3D interaction motions in CHI3D with textual descriptions.

We present SPHEAR, an accurate, differentiable parametric statistical 3D human head model, enabled by a novel 3D registration method based on spherical embeddings. We shift the paradigm away from the classical Non-Rigid Registration methods, which operate under various surface priors, increasing reconstruction fidelity and minimizing required human intervention. Additionally, SPHEAR is a complete model that allows not only to sample diverse synthetic head shapes and facial expressions, but also gaze directions, high-resolution color textures, surface normal maps, and hair cuts represented in detail, as strands. SP-HEAR can be used for automatic realistic visual data generation, semantic annotation, and general reconstruction tasks. Compared to state-of-the-art approaches, our components are fast and memory efficient, and experiments support the validity of our design choices and the accuracy of registration, reconstruction and generation techniques.

We present PhoMoH [‘fo℧.mo℧’], a neural network methodology to construct generative models of photorealistic 3D geometry and appearance of human heads including hair, beards, an oral cavity, and clothing. In contrast to prior work, PhoMoH models the human head using neural fields, thus supporting complex topology. Instead of learning a head model from scratch, we propose to augment an existing expressive head model with new features. Concretely, we learn a highly detailed geometry network layered on top of a mid-resolution head model together with a detailed, local geometry-aware, and disentangled color field. Our proposed architecture allows us to learn photorealistic human head models from relatively little data. The learned generative geometry and appearance networks can be sampled individually and enable the creation of diverse and realistic human heads. Extensive experiments validate our method qualitatively and across different metrics.

We present DiffHuman, a probabilistic method for photo-realistic 3D human reconstruction from a single RGB image. Despite the ill-posed nature of this problem, most methods are deterministic and output a single solution, often resulting in a lack of geometric detail and blurriness in unseen or uncertain regions. In contrast, DiffHuman predicts a proba-bility distribution over 3D reconstructions conditioned on an input 2D image, which allows us to sample multiple detailed 3D avatars that are consistent with the image. DiffHuman is implemented as a conditional diffusion model that denoises pixel-aligned 2D observations of an underlying 3D shape representation. During inference, we may sample 3D avatars by iteratively denoising 2D renders of the predicted 3D repre-sentation. Furthermore, we introduce a generator neural net-work that approximates rendering with considerably reduced runtime (55 x speed up), resulting in a novel dual-branch diffusion framework. Our experiments show that DiffHuman can produce diverse and detailed reconstructions for the parts of the person that are unseen or uncertain in the input image, while remaining competitive with the state-of-the-art when reconstructing visible surfaces.

We introduce Structured 3D Features, a model based on a novel implicit 3D representation that pools pixel-aligned image features onto dense 3D points sampled from a parametric, statistical human mesh surface. The 3D points have associated semantics and can move freely in 3D space. This allows for optimal coverage of the person of interest, beyond just the body shape, which in turn, additionally helps modeling accessories, hair, and loose clothing. Owing to this, we present a complete 3D transformer-based attention framework which, given a single image of a person in an unconstrained pose, generates an animatable 3D reconstruction with albedo and illumination decomposition, as a result of a single end-to-end model, trained semi-supervised, and with no additional postprocessing. We show that our S3F model surpasses the previous state-of-the-art on various tasks, including monocular 3D reconstruction, as well as albedo & shading estimation. Moreover, we show that the proposed methodology allows novel view synthesis, relighting, and re-posing the reconstruction, and can naturally be extended to handle multiple input images (e.g. different views of a person, or the same view, in different poses, in video). Finally, we demonstrate the editing capabilities of our model for 3D virtual try-on applications.

We propose a new approach to learned optimization where we represent the computation of an optimizer's update step using a neural network. The parameters of the optimizer are then learned by training on a set of optimization tasks with the objective to perform minimization efficiently. Our innovation is a new neural network architecture, Optimus, for the learned optimizer inspired by the classic BFGS algorithm. As in BFGS, we estimate a preconditioning matrix as a sum of rank-one updates but use a Transformerbased neural network to predict these updates jointly with the step length and direction. In contrast to several recent learned optimization-based approaches [24, 27], our formulation allows for conditioning across the dimensions of the parameter space of the target problem while remaining applicable to optimization tasks of variable dimensionality without retraining. We demonstrate the advantages of our approach on a benchmark composed of objective functions traditionally used for the evaluation of optimization algorithms, as well as on the real world-task of physics-based visual reconstruction of articulated 3d human motion.

We present \emph{DreamHuman}, a method to generate realistic animatable 3D human avatar models entirely from textual descriptions. Recent text-to-3D methods have made considerable strides in generation, but are still lacking in important aspects. Control and often spatial resolution remain limited, existing methods produce fixed rather than 3D human models that can be placed in different poses (i.e. re-posable or animatable), and anthropometric consistency for complex structures like people remains a challenge. \emph{DreamHuman} connects large text-to-image synthesis models, neural radiance fields, and statistical human body models in a novel optimization framework. This makes it possible to generate dynamic 3D human avatars with high-quality textures and learnt per-instance rigid and non rigid geometric deformations. We demonstrate that our method is capable to generate a wide variety of animatable, realistic 3D human models from text. These have diverse appearance, clothing, skin tones and body shapes, and outperform both generic text-to-3D approaches and previous text-based 3D avatar generators in visual fidelity.

We present PHORHUM, a novel, end-to-end trainable, deep neural network methodology for photorealistic 3D human reconstruction given just a monocular RGB image. Our pixel-aligned method estimates detailed 3D geometry and, for the first time, the unshaded surface color together with the scene illumination. Observing that 3D supervision alone is not sufficient for high fidelity color reconstruction, we introduce patch-based rendering losses that enable reliable color reconstruction on visible parts of the human, and detailed and plausible color estimation for the non-visible parts. Moreover, our method specifically addresses methodological and practical limitations of prior work in terms of representing geometry, albedo, and illumination effects, in an end-to-end model where factors can be effectively disentangled. In extensive experiments, we demonstrate the versatility and robustness of our approach. Our state-of-the-art results validate the method qualitatively and for different metrics, for both geometric and color reconstruction.

Modelling interactions between humans and objects in natural environments is central to many applications including gaming, virtual and mixed reality, as well as human behavior analysis and human-robot collaboration. This challenging operation scenario requires generalization to vast number of objects, scenes, and human actions. Unfortunately, there exist no such dataset. Moreover, this data needs to be acquired in diverse natural environments, which rules out 4D scanners and marker based capture systems. We present BEHAVE dataset, the first full body human-object interaction dataset with multi-view RGBD frames and corresponding 3D SMPL and object fits along with the annotated contacts between them. We record ~15k frames at 5 locations with 8 subjects performing a wide range of interactions with 20 common objects. We use this data to learn a model that can jointly track humans and objects in natural environments with an easy-to-use portable multi-camera setup. Our key insight is to predict correspondences from the human and the object to a statistical body model to obtain human-object contacts during interactions. Our approach can record and track not just the humans and objects but also their interactions, modeled as surface contacts, in 3D. Our code and data can be found at: http://virtualhumans.mpi-inf.mpg.de/behave.

We introduce DiffPhy, a differentiable physics-based model for articulated 3d human motion reconstruction from video. Applications of physics-based reasoning in human motion analysis have so far been limited, both by the complexity of constructing adequate physical models of articulated human motion, and by the formidable challenges of performing stable and efficient inference with physics in the loop. We jointly address such modeling and inference challenges by proposing an approach that combines a physically plausible body representation with anatomical joint limits, a differentiable physics simulator, and optimization techniques that ensure good performance and robustness to suboptimal local optima. In contrast to several recent methods [39], [42], [55], our approach readily supports full-body contact including interactions with objects in the scene. Most importantly, our model connects end-to-end with images, thus supporting direct gradient-based physics optimization by means of image-based loss functions. We validate the model by demonstrating that it can accurately reconstruct physically plausible 3d human motion from monocular video, both on public benchmarks with available 3d ground-truth, and on videos from the internet.

We focus on the task of estimating a physically plausi-ble articulated human motion from monocular video. Ex-isting approaches that do not consider physics often pro-duce temporally inconsistent output with motion artifacts, while state-of-the-art physics-based approaches have either been shown to work only in controlled laboratory conditions or consider simplified body-ground contact limited to feet. This paper explores how these shortcomings can be addressed by directly incorporating a fully-featured physics engine into the pose estimation process. Given an uncon-trolled, real-world scene as input, our approach estimates the ground-plane location and the dimensions of the physi-cal body model. It then recovers the physical motion by per-forming trajectory optimization. The advantage of our for-mulation is that it readily generalizes to a variety of scenes that might have diverse ground properties and supports any form of self-contact and contact between the articu-lated body and scene geometry. We show that our approach achieves competitive results with respect to existing physics-based methods on the Human3.6M benchmark [13], while being directly applicable without re-training to more complex dynamic motions from the AIST benchmark [36] and to uncontrolled internet videos.

Monocular estimation of three dimensional human self-contact is fundamental for detailed scene analysis including body language understanding and behaviour modeling. Existing 3d reconstruction methods do not focus on body regions in self-contact and consequently recover configurations that are either far from each other or self-intersecting, when they should just touch. This leads to perceptually incorrect estimates and limits impact in those very fine-grained analysis domains where detailed 3d models are expected to play an important role. To address such challenges we detect self-contact and design 3d losses to explicitly enforce it. Specifically, we develop a model for Self-Contact Prediction (SCP), that estimates the body surface signature of self-contact, leveraging the localization of self-contact in the image, during both training and inference. We collect two large datasets to support learning and evaluation: (1) HumanSC3D, an accurate 3d motion capture repository containing 1,032 sequences with 5,058 contact events and 1,246,487 ground truth 3d poses synchronized with images collected from multiple views, and (2) FlickrSC3D, a repository of 3,969 images, containing 25,297 surface-to-surface correspondences with annotated image spatial support. We also illustrate how more expressive 3d reconstructions can be recovered under self-contact signature constraints and present monocular detection of face-touch as one of the multiple applications made possible by more accurate self-contact models.

We study the task of embodied visual active learning, where an agent is set to explore a 3d environment with the goal to acquire visual scene understanding by actively selecting views for which to request annotation. While accurate on some benchmarks, today's deep visual recognition pipelines tend to not generalize well in certain real-world scenarios, or for unusual viewpoints. Robotic perception, in turn, requires the capability to refine the recognition capabilities for the conditions where the mobile system operates, including cluttered indoor environments or poor illumination. This motivates the proposed task, where an agent is placed in a novel environment with the objective of improving its visual recognition capability. To study embodied visual active learning, we develop a battery of agents - both learnt and pre-specified - and with different levels of knowledge of the environment. The agents are equipped with a semantic segmentation network and seek to acquire informative views, move and explore in order to propagate annotations in the neighbourhood of those views, then refine the underlying segmentation network by online retraining. The trainable method uses deep reinforcement learning with a reward function that balances two competing objectives: task performance, represented as visual recognition accuracy, which requires exploring the environment, and the necessary amount of annotated data requested during active exploration. We extensively evaluate the proposed models using the photorealistic Matterport3D simulator and show that a fully learnt method outperforms comparable pre-specified counterparts, even when requesting fewer annotations.

I went to the gym today, but how well did I do? And where should I improve? Ah, my back hurts slightly... User engagement can be sustained and injuries avoided by being able to reconstruct 3d human pose and motion, relate it to good training practices, identify errors, and provide early, real-time feedback. In this paper we introduce the first automatic system, AIFit, that performs 3d human sensing for fitness training. The system can be used at home, outdoors, or at the gym. AIFit is able to reconstruct 3d human pose, shape, and motion, reliably segment exercise repetitions, and identify in real-time the deviations between standards learnt from trainers, and the execution of a trainee. As a result, localized, quantitative feedback for correct execution of exercises, reduced risk of injury, and continuous improvement is possible. To support research and evaluation, we introduce the first large scale dataset, Fit3D, containing over 3 million images and corresponding 3d human shape and motion capture ground truth configurations, with over 37 repeated exercises, covering all the major muscle groups, performed by instructors and trainees. Our statistical coach is governed by a global parameter that captures how critical it should be of a trainee’s performance. This is an important aspect that helps adapt to a student’s level of fitness (i.e. beginner vs. advanced vs. expert), or to the expected accuracy of a 3d pose reconstruction method. We show that, for different values of the global parameter, our feedback system based on 3d pose estimates achieves good accuracy compared to the one based on ground-truth motion capture. Our statistical coach offers feedback in natural language, and with spatio-temporal visual grounding.

The detection of 3D objects from LiDAR data is a critical component in most autonomous driving systems. Safe, high speed driving needs larger detection ranges, which are enabled by new LiDARs. These larger detection ranges require more efficient and accurate detection models. Towards this goal, we propose Range Sparse Net (RSN) – a simple, efficient, and accurate 3D object detector – in order to tackle real time 3D object detection in this extended detection regime. RSN predicts foreground points from range images and applies sparse convolutions on the selected foreground points to detect objects. The lightweight 2D convolutions on dense range images results in significantly fewer selected foreground points, thus enabling the later sparse convolutions in RSN to efficiently operate. Combining features from the range image further enhance detection accuracy. RSN runs at more than 60 frames per second on a 150m × 150m detection region on Waymo Open Dataset (WOD) while being more accurate than previously published detectors. As of 11/2020, RSN is ranked first in the WOD leaderboard based on the APH/LEVEL_1 metrics for LiDAR-based pedestrian and vehicle detection, while being several times faster than alternatives.

We present deep neural network methodology to reconstruct the 3d pose and shape of people, including hand gestures and facial expression, given an input RGB image. We rely on a recently introduced, expressive full body statistical 3d human model, GHUM, trained end-to-end, and learn to reconstruct its pose and shape state in a self-supervised regime. Central to our methodology, is a learning to learn and optimize approach, referred to as HUman Neural Descent (HUND), which avoids both second-order differentiation when training the model parameters, and expensive state gradient descent in order to accurately minimize a semantic differentiable rendering loss at test time. Instead, we rely on novel recurrent stages to update the pose and shape parameters such that not only losses are minimized effectively, but the process is meta-regularized in order to ensure endprogress. HUND’s symmetry between training and testing makes it the first 3d human sensing architecture to natively support different operating regimes including self-supervised ones. In diverse tests, we show that HUND achieves very competitive results in datasets like H3.6M and 3DPW, as well as good quality 3d reconstructions for complex imagery collected in-the-wild.

We present imGHUM, the first holistic generative model of 3D human shape and articulated pose, represented as a signed distance function. In contrast to prior work, we model the full human body implicitly as a function zero-level-set and without the use of an explicit template mesh. We propose a novel network architecture and a learning paradigm, which make it possible to learn a detailed implicit generative model of human pose, shape, and semantics, on par with state-of-the-art mesh-based models. Our model features desired detail for human models, such as articulated pose including hand motion and facial expressions, a broad spectrum of shape variations, and can be queried at arbitrary resolutions and spatial locations. Additionally, our model has attached spatial semantics making it straightforward to establish correspondences between different shape instances, thus enabling applications that are difficult to tackle using classical implicit representations. In extensive experiments, we demonstrate the model accuracy and its applicability to current research problems.

We present THUNDR, a transformer-based deep neural network methodology to reconstruct the 3d pose and shape of people, given monocular RGB images. Key to our methodology is an intermediate 3d marker representation, where we aim to combine the predictive power of model-free-output architectures and the regularizing, anthropometrically-preserving properties of a statistical human surface model like GHUM—a recently introduced, expressive full body statistical 3d human model, trained end-to-end. Our novel transformer-based prediction pipeline can focus on image regions relevant to the task, supports self-supervised regimes, and ensures that solutions are consistent with human anthropometry. We show state-of-the-art results on Human3.6M and 3DPW, for both the fully-supervised and the self-supervised models, for the task of inferring 3d human shape, joint positions, and global translation. Moreover, we observe very solid 3d reconstruction performance for difficult human poses collected in the wild.

The three-dimensional reconstruction of multiple interacting humans given a monocular image is crucial for the general task of scene understanding, as capturing the subtleties of interaction is often the very reason for taking a picture. Current 3D human reconstruction methods either treat each person independently, ignoring most of the context, or reconstruct people jointly, but cannot recover interactions correctly when people are in close proximity. In this work, we introduce \textbf{REMIPS}, a model for 3D \underline{Re}construction of \underline{M}ultiple \underline{I}nteracting \underline{P}eople under Weak \underline{S}upervision. \textbf{REMIPS} can reconstruct a variable number of people directly from monocular images. At the core of our methodology stands a novel transformer network that combines unordered person tokens (one for each detected human) with positional-encoded tokens from image features patches. We introduce a novel unified model for self- and interpenetration-collisions based on a mesh approximation computed by applying decimation operators. We rely on self-supervised losses for flexibility and generalisation in-the-wild and incorporate self-contact and interaction-contact losses directly into the learning process. With \textbf{REMIPS}, we report state-of-the-art quantitative results on common benchmarks even in cases where no 3D supervision is used. Additionally, qualitative visual results show that our reconstructions are plausible in terms of pose and shape and coherent for challenging images, collected in-the-wild, where people are often interacting.

Flatness of the loss curve is conjectured to be connected to the generalization ability of machine learning models, in particular neural networks. While it has been empirically observed that flatness measures consistently correlate strongly with generalization, it is still an open theoretical problem why and under which circumstances flatness is connected to generalization, in particular in light of reparameterizations that change certain flatness measures but leave generalization unchanged. We investigate the connection between flatness and generalization by relating it to the interpolation from representative data, deriving notions of representativeness, and feature robustness. The notions allow us to rigorously connect flatness and generalization and to identify conditions under which the connection holds. Moreover, they give rise to a novel, but natural relative flatness measure that correlates strongly with generalization, simplifies to ridge regression for ordinary least squares, and solves the reparameterization issue.

We present neural radiance fields for rendering and temporal (4D) reconstruction of humans in motion (H-NeRF), as captured by a sparse set of cameras or even from a monocular video. Our approach combines ideas from neural scene representation, novel-view synthesis, and implicit statistical geometric human representations, coupled using novel loss functions. Instead of learning a radiance field with a uniform occupancy prior, we constrain it by a structured implicit human body model, represented using signed distance functions. This allows us to robustly fuse information from sparse views and generalize well beyond the poses or views observed in training. Moreover, we apply geometric constraints to co-learn the structure of the observed subject -- including both body and clothing -- and to regularize the radiance field to geometrically plausible solutions. Extensive experiments on multiple datasets demonstrate the robustness and the accuracy of our approach, its generalization capabilities significantly outside a small training set of poses and views, and statistical extrapolation beyond the observed shape.

Understanding the loss surface of neural networks is essential for the design of models with predictable performance and their success in applications. Experimental results suggest that sufficiently deep and wide neural networks are not negatively impacted by suboptimal local minima. Despite recent progress, the reason for this outcome is not fully understood. Could deep networks have very few, if at all, suboptimal local optima? or could all of them be equally good? We provide a construction to show that suboptimal local minima (i.e., non-global ones), even though degenerate, exist for fully connected neural networks with sigmoid activation functions. The local minima obtained by our construction belong to a connected set of local solutions that can be escaped from via a non-increasing path on the loss curve. For extremely wide neural networks of decreasing width after the wide layer, we prove that every suboptimal local minimum belongs to such a connected set. This provides a partial explanation for the successful application of deep neural networks. In addition, we also characterize under what conditions the same construction leads to saddle points instead of local minima for deep neural networks.

Predictability is important to autistic individuals, and robots have been suggested to meet this need as they can be programmed to be predictable, as well as elicit social interaction. The effectiveness of robot-assisted interventions designed for social skill learning presumably depends on the interplay between robot predictability, engagement in learning, and the individual differences between different autistic children. To better understand this interplay, we report on a study where 24 autistic children participated in a robot-assisted intervention. We manipulated the variance in the robot’s behaviour as a way to vary predictability, and measured the children’s behavioural engagement, visual attention, as well as their individual factors. We found that the children will continue engaging in the activity behaviourally, but may start to pay less visual attention over time to activity-relevant locations when the robot is less predictable. Instead, they increasingly start to look away from the activity. Ultimately, this could negatively influence learning, in particular for tasks with a visual component. Furthermore, severity of autistic features and expressive language ability had a significant impact on behavioural engagement. We consider our results as preliminary evidence that robot predictability is an important factor for keeping children in a state where learning can occur.

Most 3d human pose estimation methods assume that input – be it images of a scene collected from one or several viewpoints, or from a video – is given. Consequently, they focus on estimates leveraging prior knowledge and measurement by fusing information spatially and/or temporally, whenever available. In this paper we address the problem of an active observer with freedom to move and explore the scene spatially – in ‘time-freeze’ mode – and/or temporally, by selecting informative viewpoints that improve its estimation accuracy. Towards this end, we introduce Pose-DRL, a fully trainable deep reinforcement learning-based active pose estimation architecture which learns to select appropriate views, in space and time, to feed an underlying monocular pose estimator. We evaluate our model using single- and multi-target estimators with strong result in both settings. Our system further learns automatic stopping conditions in time and transition functions to the next temporal processing step in videos. In extensive experiments with the Panoptic multi-view setup, and for complex scenes containing multiple people, we show that our model learns to select viewpoints that yield significantly more accurate pose estimates compared to strong multi-view baselines.

Generating good quality and geometrically plausible synthetic images of humans with the ability to control appearance, pose and shape parameters, has become increasingly important for a variety of tasks ranging from photo editing, fashion virtual try-on, to special effects and image compression. In this paper, we propose a HUSC (HUman Synthesis and Scene Compositing) framework for the realistic synthesis of humans with different appearance, in novel poses and scenes. Central to our formulation is 3d reasoning for both people and scenes, in order to produce realistic collages, by correctly modeling perspective effects and occlusion, by taking into account scene semantics and by adequately handling relative scales. Conceptually our framework consists of three components: (1) a human image synthesis model with controllable pose and appearance, based on a parametric representation, (2) a person insertion procedure that leverages the geometry and semantics of the 3d scene, and (3) an appearance compositing process to create a seamless blending between the colors of the scene and the generated human image, and avoid visual artifacts. The performance of our framework is supported by both qualitative and quantitative results, in particular state-of-the art synthesis scores for the DeepFashion dataset.

Understanding 3d human interactions is fundamental for fine grained scene analysis and behavioural modeling. However, most of the existing models focus on analyzing a single person in isolation, and those who process several people focus largely on resolving multi-person data association, rather than inferring interactions. This may lead to incorrect, lifeless 3d estimates, that miss the subtle human contact aspects--the essence of the event--and are of little use for detailed behavioral understanding. This paper addresses such issues and makes several contributions: (1) we introduce models for interaction signature estimation (ISP) encompassing contact detection, segmentation, and 3d contact signature prediction; (2) we show how such components can be leveraged in order to produce augmented losses that ensure contact consistency during 3d reconstruction; (3) we construct several large datasets for learning and evaluating 3d contact prediction and reconstruction methods; specifically, we introduce CHI3D, a lab-based accurate 3d motion capture dataset with 631 sequences containing 2,525 contact events, 728,664 ground truth 3d poses, as well as FlickrCI3D, a dataset of 11,216 images, with 14,081 processed pairs of people, and 81,233 facet-level surface correspondences within 138,213 selected contact regions. Finally, (4) we present models and baselines to illustrate how contact estimation supports meaningful 3d reconstruction where essential interactions are captured. Models and data are made available for research purposes at http://vision.imar.ro/ci3d.

We present a statistical, articulated 3D human shape modeling pipeline, within a fully trainable, modular, deep learning framework. Given high-resolution complete 3D body scans of humans, captured in various poses, together with additional closeups of their head and facial expressions, as well as hand articulation, and given initial, artist designed, gender neutral rigged quad-meshes, we train all model parameters including non-linear shape spaces based on variational auto-encoders, pose-space deformation correctives, skeleton joint center predictors, and blend skinning functions, in a single consistent learning loop. The models are simultaneously trained with all the 3d dynamic scan data (over 60,000 diverse human configurations in our new dataset) in order to capture correlations and ensure consistency of various components. Models support facial expression analysis, as well as body (with detailed hand) shape and pose estimation. We provide fully train-able generic human models of different resolutions- the moderate-resolution GHUM consisting of 10,168 vertices and the low-resolution GHUML(ite) of 3,194 vertices-, run comparisons between them, analyze the impact of different components and illustrate their reconstruction from image data. The models will be available for research.

We address the problem of fitting 3D human models to 3D scans of dressed humans. Classical methods optimize both the data-to-model correspondences and the human model parameters (pose and shape), but are reliable only when initialised close to the solution. Some methods initialize the optimization based on fully supervised correspondence predictors, which is not differentiable end-to-end, and can only process a single scan at a time. Our main contribution is LoopReg, an end-to-end learning framework to register a corpus of scans to a common 3D human model. The key idea is to create a self-supervised loop. A backward map, parameterized by a Neural Network, predicts the correspondence from every scan point to the surface of the human model. A forward map, parameterized by a human model, transforms the corresponding points back to the scan based on the model parameters (pose and shape), thus closing the loop. Formulating this closed loop is not straightforward because it is not trivial to force the output of the NN to be on the surface of the human model -- outside this surface the human model is not even defined. To this end, we propose two key innovations. First, we define the canonical surface implicitly as the zero level set of a distance field in R3, which in contrast to more common UV parameterizations does not require cutting the surface, does not have discontinuities, and does not induce distortion. Second, we diffuse the human model to the 3D domain. This allows to map the NN predictions forward, even when they slightly deviate from the zero level set. Results demonstrate that we can train LoopReg mainly self-supervised -- following a supervised warm-start, the model becomes increasingly more accurate as additional unlabelled raw scans are processed. Our code and pre-trained models can be downloaded for research.

We present GLNet, a self-supervised framework for learning depth, optical flow, camera pose and intrinsic parameters from monocular video - addressing the difficulty of acquiring realistic ground-truth for such tasks. We propose three contributions: 1) we design new loss functions that capture multiple geometric constraints (eg. epipolar geometry) as well as adaptive photometric loss that supports multiple moving objects, rigid and non-rigid, 2) we extend the model such that it predicts camera intrinsics, making it applicable to uncalibrated video, and 3) we propose several online refinement strategies that rely on the symmetry of our self-supervised loss in training and testing, in particular optimizing model parameters and/or the output of different tasks, leveraging their mutual interactions. The idea of jointly optimizing the system output, under all geometric and photometric constraints can be viewed as a dense generalization of classical bundle adjustment. We demonstrate the effectiveness of our method on KITTI and Cityscapes, where we outperform previous self-supervised approaches on multiple tasks. We also show good generalization for transfer learning.

Existing state-of-the-art estimation systems can detect 2d poses of multiple people in images quite reliably. In contrast, 3d pose estimation from a single image is ill-posed due to occlusion and depth ambiguities. Assuming access to multiple cameras, or given an active system able to position itself to observe the scene from multiple viewpoints, reconstructing 3d pose from 2d measurements becomes well-posed within the framework of standard multi-view geometry. Less clear is what is an informative set of viewpoints for accurate 3d reconstruction, particularly in complex scenes, where people are occluded by others or by scene objects. In order to address the view selection problem in a principled way, we here introduce ACTOR, an active triangulation agent for 3d human pose reconstruction. Our fully trainable agent consists of a 2d pose estimation network (any of which would work) and a deep reinforcement learning-based policy for camera viewpoint selection. The policy predicts observation viewpoints, the number of which varies adaptively depending on scene content, and the associated images are fed to an underlying pose estimator. Importantly, training the policy requires no annotations - given a 2d pose estimator, ACTOR is trained in a self-supervised manner. In extensive evaluations on complex multi-people scenes filmed in a Panoptic dome, under multiple viewpoints, we compare our active triangulation agent to strong multi-view baselines, and show that ACTOR produces significantly more accurate 3d pose reconstructions. We also provide a proof-of-concept experiment indicating the potential of connecting our view selection policy to a physical drone observer.

We introduce new, fine-grained action and emotion recognition tasks defined on non-staged videos, recorded during robot-assisted therapy sessions of children with autism. The tasks present several challenges: a large dataset with long videos, a large number of highly variable actions, children that are only partially visible, have different ages and may show unpredictable behaviour, as well as non-standard camera viewpoints. We investigate how state-of-the-art 3d human pose reconstruction methods perform on the newly introduced tasks and propose extensions to adapt them to deal with these challenges. We also analyze multiple approaches in action and emotion recognition from 3d human pose data, establish several baselines, and discuss results and their implications in the context of child-robot interaction.

Semantic video segmentation is challenging due to the sheer amount of data that needs to be processed and labeled in order to construct accurate models. In this paper we present a deep, end-to-end trainable methodology for video segmentation that is capable of leveraging the information present in unlabeled data, besides sparsely labeled frames, in order to improve semantic estimates. Our model combines a convolutional architecture and a spatiotemporal transformer recurrent layer that is able to temporally propagate labeling information by means of optical flow, adaptively gated based on its locally estimated uncertainty. The flow, the recognition and the gated temporal propagation modules can be trained jointly, end-to-end. The temporal, gated recurrent flow propagation component of our model can be plugged into any static semantic segmentation architecture and turn it into a weakly supervised video processing one. Our experiments in the challenging CityScapes and Camvid datasets, and for multiple deep architectures, indicate that the resulting model can leverage unlabeled temporal frames, next to a labeled one, in order to improve both the video segmentation accuracy and the consistency of its temporal labeling, at no additional annotation cost and with little extra computation.

We propose drl-RPN, a deep reinforcement learning-based visual recognition model consisting of a sequential region proposal network (RPN) and an object detector. In contrast to typical RPNs, where candidate object regions (RoIs) are selected greedily via class-agnostic NMS, drl-RPN optimizes an objective closer to the final detection task. This is achieved by replacing the greedy RoI selection process with a sequential attention mechanism which is trained via deep reinforcement learning (RL). Our model is capable of accumulating class-specific evidence over time, potentially affecting subsequent proposals and classification scores, and we show that such context integration significantly boosts detection accuracy. Moreover, drl-RPN automatically decides when to stop the search process and has the benefit of being able to jointly learn the parameters of the policy and the detector, both represented as deep networks. Our model can further learn to search over a wide range of exploration-accuracy trade-offs making it possible to specify or adapt the exploration extent at test time. The resulting search trajectories are image- and category-dependent, yet rely only on a single policy over all object categories. Results on the MS COCO and PASCAL VOC challenges show that our approach outperforms established, typical state-of-the-art object detection pipelines.

Human sensing has greatly benefited from recent advances in deep learning, parametric human modeling, and large scale 2d and 3d datasets. However, existing 3d models make strong assumptions about the scene, considering either a single person per image, full views of the person, a simple background or many cameras. In this paper, we leverage state-of-the-art deep multi-task neural networks and parametric human and scene modeling, towards a fully automatic monocular visual sensing system for multiple interacting people, which (i) infers the 2d and 3d pose and shape of multiple people from a single image, relying on detailed semantic representations at both model and image level, to guide a combined optimization with feedforward and feedback components, (ii) automatically integrates scene constraints including ground plane support and simultaneous volume occupancy by multiple people, and (iii) extends the single image model to video by optimally solving the temporal person assignment problem and imposing coherent temporal pose and motion reconstructions while preserving image alignment fidelity. We perform experiments on both single and multi-person datasets, and systematically evaluate each component of the model, showing improved performance and extensive multiple human sensing capability. We also apply our method to images with multiple people, severe occlusions and diverse backgrounds captured in challenging natural scenes, and obtain results of good perceptual quality.

We propose an automatic person-to-person appearance transfer model based on explicit parametric 3d human representations and learned, constrained deep translation network architectures for photographic image synthesis. Given a single source image and a single target image, each corresponding to different human subjects, wearing different clothing and in different poses, our goal is to photo-realistically transfer the appearance from the source image onto the target image while preserving the target shape and clothing segmentation layout. Our solution to this new problem is formulated in terms of a computational pipeline that combines (1) 3d human pose and body shape estimation from monocular images, (2) identifying 3d surface colors elements (mesh triangles) visible in both images, that can be transferred directly using barycentric procedures, and (3) predicting surface appearance missing in the first image but visible in the second one using deep learning-based image synthesis techniques. Our model achieves promising results as supported by a perceptual user study where the participants rated around 65% of our results as good, very good or perfect, as well in automated tests (Inception scores and a Faster-RCNN human detector responding very similarly to real and model generated images). We further show how the proposed architecture can be profiled to automatically generate images of a person dressed with different clothing transferred from a person in another image, opening paths for applications in entertainment and photo-editing (e.g. embodying and posing as friends or famous actors), the fashion industry, or affordable online shopping of clothing.

The problem of graph matching under node and pairwise constraints is fundamental in areas as diverse as combinatorial optimization, machine learning or computer vision, where representing both the relations between nodes and their neighborhood structure is essential. We present an end-to-end model that makes it possible to learn all parameters of the graph matching process, including the unary and pairwise node neighborhoods, represented as deep feature extraction hierarchies. The challenge is in the formulation of the different matrix computation layers of the model in a way that enables the consistent, efficient propagation of gradients in the complete pipeline from the loss function, through the combinatorial optimization layer solving the matching problem, and the feature extraction hierarchy. Our computer vision experiments and ablation studies on challenging datasets like PASCAL VOC keypoints, Sintel and CUB show that matching models refined end-to-end are superior to counterparts based on feature hierarchies trained for other problems.

We present MubyNet -- a feed-forward, multitask, bottom up system for the integrated localization, as well as 3d pose and shape estimation, of multiple people in monocular images. The challenge is the formal modeling of the problem that intrinsically requires discrete and continuous computation, e.g. grouping people vs. predicting 3d pose. The model identifies human body structures (joints and limbs) in images, groups them based on 2d and 3d information fused using learned scoring functions, and optimally aggregates such responses into partial or complete 3d human skeleton hypotheses under kinematic tree constraints, but without knowing in advance the number of people in the scene and their visibility relations. We design a multi-task deep neural network with differentiable stages where the person grouping problem is formulated as an integer program based on learned body part scores parameterized by both 2d and 3d information. This avoids suboptimality resulting from separate 2d and 3d reasoning, with grouping performed based on the combined representation. The final stage of 3d pose and shape prediction is based on a learned attention process where information from different human body parts is optimally integrated. State-of-the-art results are obtained in large scale datasets like Human3.6M and Panoptic, and qualitatively by reconstructing the 3d shape and pose of multiple people, under occlusion, in difficult monocular images.

Learning a computationally efficient kernel from data is an important machine learning problem. The majority of kernels in the literature do not leverage the geometry of the data, and those that do are computationally infeasible for contemporary datasets. Recent advances in approximation techniques have expanded the applicability of the kernel methodology to scale linearly with the data size. Data-dependent kernels, which could leverage this computational advantage, have however not yet seen the benefit. Here we derive an approximate large-scale learning procedure for data-dependent kernels that is efficient and performs well in practice. We provide a Lemma that can be used to derive the asymptotic convergence of the approximation in the limit of infinite random features, and, under certain conditions, an estimate of the convergence speed. We empirically prove that our construction represents a valid, yet efficient approximation of the data-dependent kernel. For large-scale datasets of millions of datapoints, where the proposed method is now applicable for the first time, we notice a significant performance boost over both baselines consisting of data independent kernels and of kernel approximations, at comparable computational cost.

We propose a deep multitask architecture for fully automatic 2d and 3d human sensing (DMHS), including recognition and reconstruction, in monocular images. The system computes the figure-ground segmentation, semantically identifies the human body parts at pixel level, and estimates the 2d and 3d pose of the person. The model supports the joint training of all components by means of multi-task losses where early processing stages recursively feed into advanced ones for increasingly complex calculations, accuracy and robustness. The design allows us to tie a complete training protocol, by taking advantage of multiple datasets that would otherwise restrictively cover only some of the model components: complex 2d image data with no body part labeling and without associated 3d ground truth, or complex 3d data with limited 2d background variability. In detailed experiments based on several challenging 2d and 3d datasets (LSP, HumanEva, Human3.6M), we evaluate the sub-structures of the model, the effect of various types of training data in the multitask loss, and demonstrate that state-of-the-art results can be achieved at all processing levels. We also show that in the wild our monocular RGB architecture is perceptually competitive to a state-of-the art (commercial) Kinect system based on RGB-D data.

One of the most widely used strategies for visual object detection is based on exhaustive spatial hypothesis search. While methods like sliding windows have been successful and effective for many years, they are still brute-force, independent of the image content and the visual category being searched. In this paper we present principled sequential models that accumulate evidence collected at a small set of image locations in order to detect visual objects effectively. By formulating sequential search as reinforcement learning of the search policy (including the stopping condition), our fully trainable model can explicitly balance for each class, specifically, the conflicting goals of exploration - sampling more image regions for better accuracy -, and exploitation - stopping the search efficiently when sufficiently confident about the target's location. The methodology is general and applicable to any detector response function. We report encouraging results in the PASCAL VOC 2012 object detection test set showing that the proposed methodology achieves almost two orders of magnitude speed-up over sliding window methods.

We focus on the problem of semantic segmentation based on RGB-D data, with emphasis on analyzing cluttered indoor scenes containing many visual categories and instances. Our approach is based on a parametric figureground intensity and depth-constrained proposal process that generates spatial layout hypotheses at multiple locations and scales in the image followed by a sequential inference algorithm that produces a complete scene estimate. Our contributions can be summarized as follows: (1) a generalization of parametric max flow figure-ground proposal methodology to take advantage of intensity and depth information, in order to systematically and efficiently generate the breakpoints of an underlying spatial model in polynomial time, (2) new region description methods based on second-order pooling over multiple features constructed using both intensity and depth channels, (3) a principled search-based structured prediction inference and learning process that resolves conflicts in overlapping spatial partitions and selects regions sequentially towards complete scene estimates, and (4) extensive evaluation of the impact of depth, as well as the effectiveness of a large number of descriptors, both pre-designed and automatically obtained using deep learning, in a difficult RGB-D semantic segmentation problem with 92 classes. We report state of the art results in the challenging NYU Depth Dataset V2 [44], extended for the RMRC 2013 and RMRC 2014 Indoor Segmentation Challenges, where currently the proposed model ranks first. Moreover, we show that by combining second-order and deep learning features, over 15% relative accuracy improvements can be additionally achieved. In a scene classification benchmark, our methodology further improves the state of the art by 24%.

Deep neural network architectures have recently produced excellent results in a variety of areas in artificial intelligence and visual recognition, well surpassing traditional shallow architectures trained using hand-designed features. The power of deep networks stems both from their ability to perform local computations followed by pointwise non-linearities over increasingly larger receptive fields, and from the simplicity and scalability of the gradient-descent training procedure based on backpropagation. An open problem is the inclusion of layers that perform global, structured matrix computations like segmentation (e.g. normalized cuts) or higher-order pooling (e.g. log-tangent space metrics defined over the manifold of symmetric positive definite matrices) while preserving the validity and efficiency of an end-to-end deep training framework. In this paper we propose a sound mathematical apparatus to formally integrate global structured computation into deep computation architectures. At the heart of our methodology is the development of the theory and practice of backpropagation that generalizes to the calculus of adjoint matrix variations. We perform segmentation experiments using the BSDS and MSCOCO benchmarks and demonstrate that deep networks relying on second-order pooling and normalized cuts layers, trained end-to-end using matrix backpropagation, outperform counterparts that do not take advantage of such global layers.

The emergence of modern, affordable and accurate RGB-D sensors increases the need for single view approaches to estimate 3-dimensional motion, also known as scene flow. In this paper we propose a coarse-to-fine, dense, correspondence-based scene flow formulation that relies on explicit geometric reasoning to account for the effects of large displacements and to model occlusion. Our methodology enforces local motion rigidity at the level of the 3d point cloud without explicitly smoothing the parameters of adjacent neighborhoods. By integrating all geometric and photometric components in a single, consistent, occlusion-aware energy model, defined over overlapping, image-adaptive neighborhoods, our method can process fast motions and large occlusions areas, as present in challenging datasets like the MPI Sintel Flow Dataset, recently augmented with depth information. By explicitly modeling large displacements and occlusion, we can handle difficult sequences which cannot be currently processed by state of the art scene flow methods. We also show that by integrating depth information into the model, we can obtain correspondence fields with improved spatial support and sharper boundaries compared to the state of the art, large-displacement optical flow methods.

Semantic segmentation and object detection are nowadays dominated by methods operating on regions obtained as a result of a bottom-up grouping process (segmentation) but use feature extractors developed for recognition on fixed-form (e.g. rectangular) patches, with full images as a special case. This is most likely suboptimal. In this paper we focus on feature extraction and description over free-form regions and study the relationship with their fixed-form counterparts. Our main contributions are novel pooling techniques that capture the second-order statistics of local descriptors inside such free-form regions. We introduce second-order generalizations of average and max-pooling that together with appropriate non-linearities, derived from the mathematical structure of their embedding space, lead to state-of-the-art recognition performance in semantic segmentation experiments without any type of local feature coding. In contrast, we show that codebook-based local feature coding is more important when feature extraction is constrained to operate over regions that include both foreground and large portions of the background, as typical in image classification settings, whereas for high-accuracy localization setups, second-order pooling over free-form regions produces results superior to those of the winning systems in the contemporary semantic segmentation challenges, with models that are much faster in both training and testing.

Systems based on bag-of-words models from image features collected at maxima of sparse interest point operators have been used successfully for both computer visual object and action recognition tasks. While the sparse, interest-point based approach to recognition is not inconsistent with visual processing in biological systems that operate in `saccade and fixate' regimes, the methodology and emphasis in the human and the computer vision communities remains sharply distinct. Here, we make three contributions aiming to bridge this gap. First, we complement existing state-of-the art large scale dynamic computer vision annotated datasets like Hollywood-2 [1] and UCF Sports [2] with human eye movements collected under the ecological constraints of visual action and scene context recognition tasks. To our knowledge these are the first large human eye tracking datasets to be collected and made publicly available for video, vision.imar.ro/eyetracking (497,107 frames, each viewed by 19 subjects), unique in terms of their (a) large scale and computer vision relevance, (b) dynamic, video stimuli, (c) task control, as well as free-viewing. Second, we introduce novel dynamic consistency and alignment measures, which underline the remarkable stability of patterns of visual search among subjects. Third, we leverage the significant amount of collected data in order to pursue studies and build automatic, end-to-end trainable computer vision systems based on human eye movements. Our studies not only shed light on the differences between computer vision spatio-temporal interest point image sampling strategies and the human fixations, as well as their impact for visual recognition performance, but also demonstrate that human fixations can be accurately predicted, and when used in an end-to-end automatic system, leveraging some of the advanced computer vision practice, can lead to state of the art results.

Recently, the emergence of Kinect systems has demonstrated the benefits of predicting an intermediate body part labeling for 3D human pose estimation, in conjunction with RGB-D imagery. The availability of depth information plays a critical role, so an important question is whether a similar representation can be developed with sufficient robustness in order to estimate 3D pose from RGB images. This paper provides evidence for a positive answer, by leveraging (a) 2D human body part labeling in images, (b) second-order label-sensitive pooling over dynamically computed regions resulting from a hierarchical decomposition of the body, and (c) iterative structured-output modeling to contextualize the process based on 3D pose estimates. For robustness and generalization, we take advantage of a recent large-scale 3D human motion capture dataset, Human3.6M[18] that also has human body part labeling annotations available with images. We provide extensive experimental studies where alternative intermediate representations are compared and report a substantial 33% error reduction over competitive discriminative baselines that regress 3D human pose against global HOG features.

We introduce a new dataset, Human3.6M, of 3.6 Million accurate 3D Human poses, acquired by recording the performance of 5 female and 6 male subjects, under 4 different viewpoints, for training realistic human sensing systems and for evaluating the next generation of human pose estimation models and algorithms. Besides increasing the size of the datasets in the current state-of-the-art by several orders of magnitude, we also aim to complement such datasets with a diverse set of motions and poses encountered as part of typical human activities (taking photos, talking on the phone, posing, greeting, eating, etc.), with additional synchronized image, human motion capture, and time of flight (depth) data, and with accurate 3D body scans of all the subject actors involved. We also provide controlled mixed reality evaluation scenarios where 3D human models are animated using motion capture and inserted using correct 3D geometry, in complex real environments, viewed with moving cameras, and under occlusion. Finally, we provide a set of large-scale statistical models and detailed evaluation baselines for the dataset illustrating its diversity and the scope for improvement by future work in the research community. Our experiments show that our best large-scale model can leverage our full training set to obtain a 20% improvement in performance compared to a training set of the scale of the largest existing public dataset for this problem. Yet the potential for improvement by leveraging higher capacity, more complex models with our large dataset, is substantially vaster and should stimulate future research. The dataset together with code for the associated large-scale learning models, features, visualization tools, as well as the evaluation server, is available online at http://vision.imar.ro/human3.6m.

Boundary detection is a fundamental computer vision problem that is essential for a variety of tasks, such as contour and region segmentation, symmetry detection and object recognition and categorization. We propose a generalized formulation for boundary detection, with closed-form solution, applicable to the localization of different types of boundaries, such as object edges in natural images and occlusion boundaries from video. Our generalized boundary detection method (Gb) simultaneously combines low-level and mid-level image representations in a single eigenvalue problem and solves for the optimal continuous boundary orientation and strength. The closed-form solution to boundary detection enables our algorithm to achieve state-of-the-art results at a significantly lower computational cost than current methods. We also propose two complementary novel components that can seamlessly be combined with Gb: first, we introduce a soft-segmentation procedure that provides region input layers to our boundary detection algorithm for a significant improvement in accuracy, at negligible computational cost; second, we present an efficient method for contour grouping and reasoning, which when applied as a final post-processing stage, further increases the boundary detection performance.

In this paper we present an inference procedure for the semantic segmentation of images. Different from many CRF approaches that rely on dependencies modeled with unary and pairwise pixel or super pixel potentials, our method is entirely based on estimates of the overlap between each of a set of mid-level object segmentation proposals and the objects present in the image. We define continuous latent variables on super pixels obtained by multiple intersections of segments, then output the optimal segments from the inferred super pixel statistics. The algorithm is capable of recombine and refine initial mid-level proposals, as well as handle multiple interacting objects, even from the same class, all in a consistent joint inference framework by maximizing the composite likelihood of the underlying statistical model using an EM algorithm. In the PASCAL VOC segmentation challenge, the proposed approach obtains high accuracy and successfully handles images of complex object interactions.

Estimating a dense correspondence field between successive video frames, under large displacement, is important in many visual learning and recognition tasks. We propose a novel sparse-to-dense matching method for motion field estimation and occlusion detection. As an alternative to the current coarse-to-fine approaches from the optical flow literature, we start from the higher level of sparse matching with rich appearance and geometric constraints collected over extended neighborhoods, using an occlusion aware, locally affine model. Then, we move towards the simpler, but denser classic flow field model, with an interpolation procedure that offers a natural transition between the sparse and the dense correspondence fields. We experimentally demonstrate that our appearance features and our complex geometric constraints permit the correct motion estimation even in difficult cases of large displacements and significant appearance changes. We also propose a novel classification method for occlusion detection that works in conjunction with the sparse-to-dense matching model. We validate our approach on the newly released Sintel dataset and obtain state-of-the-art results.

Human motion analysis in images and video is a central computer vision problem. Yet, there are no studies that reveal how humans perceive other people in images and how accurate they are. In this paper we aim to unveil some of the processing-as well as the levels of accuracy-involved in the 3D perception of people from images by assessing the human performance. Our contributions are: (1) the construction of an experimental apparatus that relates perception and measurement, in particular the visual and kinematic performance with respect to 3D ground truth when the human subject is presented an image of a person in a given pose, (2) the creation of a dataset containing images, articulated 2D and 3D pose ground truth, as well as synchronized eye movement recordings of human subjects, shown a variety of human body configurations, both easy and difficult, as well as their 're-enacted' 3D poses, (3) quantitative analysis revealing the human performance in 3D pose re-enactment tasks, the degree of stability in the visual fixation patterns of human subjects, and the way it correlates with different poses. We also discuss the implications of our findings for the construction of visual human sensing systems.

Human action recognition under low observational latency is receiving a growing interest in computer vision due to rapidly developing technologies in human-robot interaction, computer gaming and surveillance. In this paper we propose a fast, simple, yet powerful non-parametric Moving Pose (MP) framework for low-latency human action and activity recognition. Central to our methodology is a moving pose descriptor that considers both pose information as well as differential quantities (speed and acceleration) of the human body joints within a short time window around the current frame. The proposed descriptor is used in conjunction with a modified kNN classifier that considers both the temporal location of a particular frame within the action sequence as well as the discrimination power of its moving pose descriptor compared to other frames in the training set. The resulting method is non-parametric and enables low-latency recognition, one-shot learning, and action detection in difficult unsegmented sequences. Moreover, the framework is real-time, scalable, and outperforms more sophisticated approaches on challenging benchmarks like MSR-Action3D or MSR-DailyActivities3D.

We organized a Grand Challenge and Workshop on Multi-Modal Gesture Recognition.

Human eye movements provide a rich source of information into the human visual processing. The complex interplay between the task and the visual stimulus is believed to determine human eye movements, yet it is not fully understood. This has precluded the development of reliable dynamic eye movement prediction systems. Our work makes three contributions towards addressing this problem. First, we complement one of the largest and most challenging static computer vision datasets, VOC 2012 Actions, with human eye movement annotations collected under the task constraints of action and context recognition. Our dataset is unique among eyetracking datasets for still images in terms of its large scale (over 1 million fixations, 9157 images), task control and action from a single image emphasis. Second, we introduce models to automatically discover areas of interest (AOI) and introduce novel dynamic consistency metrics, based on them. Our method can automatically determine the number and spatial support of the AOIs, in addition to their locations. Based on such encodings, we show that, on unconstrained read-world stimuli, task instructions have significant influence on visual behavior. Finally, we leverage our large scale dataset in conjunction with powerful machine learning techniques and computer vision features, to introduce novel dynamic eye movement prediction methods which learn task-sensitive reward functions from eye movement data and efficiently integrate these rewards to plan future saccades based on inverse optimal control. We show that the propose methodology achieves state of the art scanpath modeling results.

The random Fourier embedding methodology can be used to approximate the performance of non-linear kernel classifiers in linear time on the number of training examples. However, there still exists a non-trivial performance gap between the approximation and the nonlinear models, especially for the exponential χ2kernel, one of the most powerful models for histograms. Based on analogies with Chebyshev polynomials, we propose an asymptotically convergent analytic series of the χ2measure. The new series removes the need to use periodic approximations to the χ2function, as typical in previous methods, and improves the classification accuracy when used in the random Fourier approximation of the exponential χ2kernel. Besides, out-of-core principal component analysis (PCA) methods are introduced to reduce the dimensionality of the approximation and achieve better performance at the expense of only an additional constant factor to the time complexity. Moreover, when PCA is performed jointly on the training and unlabeled testing data, further performance improvements can be obtained. The proposed approaches are tested on the PASCAL VOC 2010 segmentation and the ImageNet ILSVR-C 2010 datasets, and shown to give statistically significant improvements over alternative approximation methods.

We present a novel framework to generate and rank plausible hypotheses for the spatial extent of objects in images using bottom-up computational processes and mid-level selection cues. The object hypotheses are represented as figure-ground segmentations, and are extracted automatically, without prior knowledge of the properties of individual object classes, by solving a sequence of Constrained Parametric Min-Cut problems (CPMC) on a regular image grid. In a subsequent step, we learn to rank the corresponding segments by training a continuous model to predict how likely they are to exhibit real-world regularities (expressed as putative overlap with ground truth) based on their mid-level region properties, then diversify the estimated overlap score using maximum marginal relevance measures. We show that this algorithm significantly outperforms the state of the art for low-level segmentation in the VOC 2009 and 2010 data sets. In our companion papers [1], [2], we show that the algorithm can be used, successfully, in a segmentation-based visual object category recognition pipeline. This architecture ranked first in the VOC2009 and VOC2010 image segmentation and labeling challenges.

We propose a mid-level statistical model for image segmentation that composes multiple figure-ground hypotheses (FG) obtained by applying constraints at different locations and scales, into larger interpretations (tilings) of the entire image. Inference is cast as optimization over sets of maximal cliques sampled from a graph connecting all non-overlapping figure-ground segment hypotheses. Potential functions over cliques combine unary, Gestalt-based figure qualities, and pairwise compatibilities among spatially neighboring segments, constrained by T-junctions and the boundary interface statistics of real scenes. Learning the model parameters is based on maximum likelihood, alternating between sampling image tilings and optimizing their potential function parameters. State of the art results are reported on the Berkeley and Stanford segmentation datasets, as well as VOC2009, where a 28% improvement was achieved.

We present an approach for automatic 3D human pose reconstruction from monocular images, based on a discriminative formulation with latent segmentation inputs. We advanced the field of structured prediction and human pose reconstruction on several fronts. First, by working with a pool of figure-ground segment hypotheses, the prediction problem is formulated in terms of combined learning and inference over segment hypotheses and 3D human articular configurations. Beside constructing tractable formulations for the combined segment selection and pose estimation problem, we propose new augmented kernels that can better encode complex dependencies between output variables. Furthermore, we provide primal linear re-formulations based on Fourier kernel approximations, in order to scale-up the non-linear latent structured prediction methodology. The proposed models are shown to be competitive in the HumanEva benchmark and are also illustrated in a clip collected from a Hollywood movie, where the model can infer human poses from monocular images captured in complex environments.

Graph and hypergraph matching are important problems in computer vision. They are successfully used in many applications requiring 2D or 3D feature matching, such as 3D reconstruction and object recognition. While graph matching is limited to using pairwise relationships, hypergraph matching permits the use of relationships between sets of features of any order. Consequently, it carries the promise to make matching more robust to changes in scale, deformations and outliers. In this paper we make two contributions. First, we present a first semi-supervised algorithm for learning the parameters that control the hypergraph matching model and demonstrate experimentally that it significantly improves the performance of current state-of-the-art methods. Second, we propose a novel efficient hypergraph matching algorithm, which outperforms the state-of-the-art, and, when used in combination with other higher-order matching algorithms, it consistently improves their performance.

We present a joint image segmentation and labeling model (JSL) which, given a bag of figure-ground segment hypotheses extracted at multiple image locations and scales, constructs a joint probability distribution over both the compatible image interpretations (tilings or image segmentations) composed from those segments, and over their labeling into categories. The process of drawing samples from the joint distribution can be interpreted as first sampling tilings, modeled as maximal cliques, from a graph connecting spatially non-overlapping segments in the bag, followed by sampling labels for those segments, conditioned on the choice of a particular tiling. We learn the segmentation and labeling parameters jointly, based on Maximum Likelihood with a novel Incremental Saddle Point estimation procedure. The partition function over tilings and labelings is increasingly more accurately approximated by including incorrect configurations that a not-yet-competent model rates probable during learning. We show that the proposed methodology matches the current state of the art in the Stanford dataset, as well as in VOC2010, where 41.7% accuracy on the test set is achieved.

We present a novel framework for generating and ranking plausible objects hypotheses in an image using bottom-up processes and mid-level cues. The object hypotheses are represented as figure-ground segmentations, and are extracted automatically, without prior knowledge about properties of individual object classes, by solving a sequence of constrained parametric min-cut problems (CPMC) on a regular image grid. We then learn to rank the object hypotheses by training a continuous model to predict how plausible the segments are, given their mid-level region properties. We show that this algorithm significantly outperforms the state of the art for low-level segmentation in the VOC09 segmentation dataset. It achieves the same average best segmentation covering as the best performing technique to date, 0.61 when using just the top 7 ranked segments, instead of the full hierarchy in. Our method achieves 0.78 average best covering using 154 segments. In a companion paper, we also show that the algorithm achieves state-of-the art results when used in a segmentation-based recognition pipeline.

We present an approach to visual object-class recognition and segmentation based on a pipeline that combines multiple, holistic figure-ground hypotheses generated in a bottom-up, object independent process. Decisions are performed based on continuous estimates of the spatial overlap between image segment hypotheses and each putative class. We differ from existing approaches not only in our seemingly unreasonable assumption that good object-level segments can be obtained in a feed-forward fashion, but also in framing recognition as a regression problem. Instead of focusing on a one-vs-all winning margin that can scramble ordering inside the non-maximum (non-winning) set, learning produces a globally consistent ranking with close ties to segment quality, hence to the extent entire object or part hypotheses spatially overlap with the ground truth. We demonstrate results beyond the current state of the art for image classification, object detection and semantic segmentation, in a number of challenging datasets including Caltech-101, ETHZ-Shape and PASCAL VOC 2009.

Multiple-Instance learning has been long known as a hard non-convex problem. In this work, we propose an approach that recasts it as a convex likelihood ratio estimation problem. Firstly, the constraint in multiple-instance learning is reformulated into a convex constraint on the likelihood ratio. Then we show that a joint estimation of a likelihood ratio function and the likelihood on training instances can be learned convexly. Theoretically, we prove a quantitative relationship between the risk estimated under the 0-1 classification loss, and under a loss function for likelihood ratio estimation. It is shown that our likelihood ratio estimation is generally a good surrogate for the 0-1 loss, and separates positive and negative instances well. However with the joint estimation it tends to underestimate the likelihood of an example to be positive. We propose to use these likelihood ratio estimates as features, and learn a linear combination on them to classify the bags. Experiments on synthetic and real datasets show the superiority of the approach.

Structured outputs such as multidimensional vectors or graphs are frequently encountered in real world pattern recognition applications such as computer vision, natural language processing or computational biology. This motivates the learning of functional dependencies between spaces with complex, interdependent inputs and outputs, as arising e.g. from images and their corresponding 3d scene representations. In this spirit, we propose a new structured learning method-Structured Output-Associative Regression (SOAR)-that models not only the input-dependency but also the self-dependency of outputs, in order to provide an output re-correlation mechanism that complements the (more standard) input-based regressive prediction. The model is simple but powerful, and, in principle, applicable in conjunction with any existing regression algorithms. SOAR can be kernelized to deal with non-linear problems and learning is efficient via primal/dual formulations not unlike ones used for kernel ridge regression or support vector regression. We demonstrate that the method outperforms weighted nearest neighbor and regression methods for the reconstruction of images of handwritten digits and for 3D human pose estimation from video in the HumanEva benchmark.

Recognizing human action in non-instrumented video is a challenging task not only because of the variability produced by general scene factors like illumination, background, occlusion or intra-class variability, but also because of subtle behavioral patterns among interacting people or between people and objects in images. To improve recognition, a system may need to use not only low-level spatio-temporal video correlations but also relational descriptors between people and objects in the scene. In this paper we present contextual scene descriptors and Bayesian multiple kernel learning methods for recognizing human action in complex non-instrumented video. Our contribution is threefold: (1) we introduce bag-of-detector scene descriptors that encode presence/absence and structural relations between object parts; (2) we derive a novel Bayesian classification method based on Gaussian processes with multiple kernel covariance functions (MKGPC), in order to automatically select and weight multiple features, both low-level and high-level, out of a large collection, in a principled way, and (3) perform large scale evaluation using a variety of features on the KTH and a recently introduced, challenging, Hollywood movie dataset. On the KTH dataset, we obtain 94.1% accuracy, the best result reported to date. On the Hollywood dataset we obtain promising results in several action classes using fewer descriptors and about 9.1% improvement in a previous benchmark test.

We present an integrated model for visual object localization and continuous state estimation in a discriminative structured prediction framework. While existing discriminative `prediction through time' methods have showed remarkable versatility for visual reconstruction and tracking problems, they tend to assume that the input is known (or the object is segmented) a condition that can rarely be accommodated in images of real scenes. Our structural Support Vector Machine (structSVM) framework offers an end-to-end training and inference framework that overcomes these limitations by consistently searching both in the space of possible inputs (effectively an efficient form of object localization) and in the space of possible structured outputs, given those inputs. We demonstrate the potential of this methodology for 3d human pose reconstruction in monocular images both in the HumanEva benchmark, where 3d ground truth is available, and qualitatively, in un-instrumented images of real scenes.

Skeletonization algorithms typically decompose an object's silhouette into a set of symmetric parts, offering a powerful representation for shape categorization. However, having access to an object's silhouette assumes correct figure-ground segmentation, leading to a disconnect with the mainstream categorization community, which attempts to recognize objects from cluttered images. In this paper, we present a novel approach to recovering and grouping the symmetric parts of an object from a cluttered scene. We begin by using a multiresolution superpixel segmentation to generate medial point hypotheses, and use a learned affinity function to perceptually group nearby medial points likely to belong to the same medial branch. In the next stage, we learn higher granularity affinity functions to group the resulting medial branches likely to belong to the same object. The resulting framework yields a skeletal approximation that's free of many of the instabilities plaguing traditional skeletons. More importantly, it doesn't require a closed contour, enabling the application of skeleton-based categorization systems to more realistic imagery

In visual recognition, the images are frequently modeled as sets of local features (bags). We show that bag of words, a common method to handle such cases, can be viewed as a special match kernel, which counts 1 if two local features fall into the same regions partitioned by visual words and 0 otherwise. Despite its simplicity, this quantization is too coarse. It is, therefore, appealing to design match kernels that more accurately measure the similarity between local features. However, it is impractical to use such kernels on large datasets due to their significant computational cost. To address this problem, we propose an efficient match kernel (EMK), which maps local features to a low dimensional feature space, average the resulting feature vectors to form a set-level feature, then apply a linear classifier. The local feature maps are learned so that their inner products preserve, to the best possible, the values of the specified kernel function. EMK is linear both in the number of images and in the number of local features. We demonstrate that EMK is extremely efficient and achieves the current state of the art performance on three difficult real world datasets: Scene-15, Caltech-101 and Caltech-256.

The potential success of discriminative learning approaches to 3D reconstruction relies on the ability to efficiently train predictive algorithms using sufficiently many examples that are representative of the typical configurations encountered in the application domain. Recent research indicates that sparse conditional Bayesian mixture of experts (cMoE) models (e.g. BME (Sminchisescu et al., 2005)) are adequate modeling tools that not only provide contextual 3D predictions for problems like human pose reconstruction, but can also represent multiple interpretations that result from depth ambiguities or occlusion. However, training conditional predictors requires sophisticated double-loop algorithms that scale unfavorably with the input dimension and the training set size, thus limiting their usage to 10,000 examples of less, so far. In this paper we present large-scale algorithms, referred to as fBME, that combine forward feature selection and bound optimization in order to train probabilistic, BME models, with one order of magnitude more data (100,000 examples and up) and more than one order of magnitude faster. We present several large scale experiments, including monocular evaluation on the HumanEva dataset (Sigal and Black, 2006), demonstrating how the proposed methods overcome the scaling limitations of existing ones.

Recent research in visual inference from monocular images has shown that discriminatively trained image-based predictors can provide fast, automatic qualitative 3D reconstructions of human body pose or scene structure in real-world environments. However, the stability of existing image representations tends to be perturbed by deformations and misalignments in the training set, which, in turn, degrade the quality of learning and generalization. In this paper we advocate the semi-supervised learning of hierarchical image descriptions in order to better tolerate variability at multiple levels of detail. We combine multilevel encodings with improved stability to geometric transformations, with metric learning and semi-supervised manifold regularization methods in order to further profile them for task-invariance -resistance to background clutter and within the same human pose class differences. We quantitatively analyze the effectiveness of both descriptors and learning methods and show that each one can contribute, sometimes substantially, to more reliable 3D human pose estimates in cluttered images.

We propose non-linear generative models referred to as Sparse Spectral Latent Variable Models (SLVM), that combine the advantages of spectral embeddings with the ones of parametric latent variable models: (1) provide stable latent spaces that preserve global or local geometric properties of the modeled data; (2) offer low-dimensional generative models with probabilistic, bi-directional mappings between latent and ambient spaces, (3) are probabilistically consistent (i.e., reflect the data distribution, both jointly and marginally) and efficient to learn and use. We show that SLVMs compare favorably with competing methods based on PCA, GPLVM or GTM for the reconstruction of typical human motions like walking, running, pantomime or dancing in a benchmark dataset. Empirically, we observe that SLVMs are effective for the automatic 3d reconstruction of low-dimensional human motion in movies.

Kernel classifiers based on Support Vector Machines (SVM) have recently achieved state-of-the art results on several popular datasets like Caltech or Pascal. This was possible by combining the advantages of SVM - convexity and the availability of efficient optimizers, with 'hyperkernels' - linear combinations of kernels computed at multiple levels of image encoding. The use of hyperkernels faces the challenge of choosing the kernel weights, the use of possibly irrelevant, poorly performing kernels, and an increased number of parameters that can lead to overfitting. In this paper we advocate the transition from SVMs to Support Kernel Machines (SKM) - models that estimate both the parameters of a sparse linear combination of kernels, and the parameters of a discriminative classifier. We exploit recent kernel learning techniques, not previously used in computer vision, that show how learning SKMs can be formulated as a convex optimization problem, which can be solved efficiently using Sequential Minimal Optimization. We study kernel learning for several multi-level image encodings for supervised object recognition and report competitive results on several datasets, including INRIA pedestrian, Caltech 101 and the newly created Caltech 256.

Reliably recovering 3D human pose from monocular video requires constraints that bias the estimates towards typical human poses and motions. We define priors for people tracking using a Laplacian Eigenmaps Latent Variable Model (LELVM). LELVM is a probabilistic dimensionality reduction model that naturally combines the advantages of latent variable models---definining a multimodal probability density for latent and observed variables, and globally differentiable nonlinear mappings for reconstruction and dimensionality reduction---with those of spectral manifold learning methods---no local optima, ability to unfold highly nonlinear manifolds, and good practical scaling to latent spaces of high dimension. LELVM is computationally efficient, simple to learn from sparse training data, and compatible with standard probabilistic trackers such as particle filters. We analyze the performance of a LELVM-based probabilistic sigma point mixture tracker in several real and synthetic human motion sequences and demonstrate that LELVM provides sufficient constraints for robust operation in the presence of missing, noisy and ambiguous image measurements.

We introduce BM3 E, a Conditional Bayesian Mixture of Experts Markov Model, for consistent probabilistic estimates in discriminative visual tracking. The model applies to problems of temporal and uncertain inference and represents the unexplored bottom-up counterpart of pervasive generative models estimated with Kalman filtering or particle filtering. Instead of inverting a non-linear generative observation model at run-time, we learn to cooperatively predict complex state distributions directly from descriptors that encode image observations - typically bag-of-feature global image histograms or descriptors computed over regular spatial grids. These are integrated in a conditional graphical model in order to enforce temporal smoothness constraints and allow a principled management of uncertainty. The algorithms combine sparsity, mixture modeling, and non-linear dimensionality reduction for efficient computation in high-dimensional continuous state spaces. The combined system automatically self-initializes and recovers from failure. The research has three contributions: (1) We establish the density propagation rules for discriminative inference in continuous, temporal chain models; (2) We propose flexible supervised and unsupervised algorithms for learning feedforward, multivalued contextual mappings (multimodal state distributions) based on compact, conditional Bayesian mixture of experts models; (3) We validate the framework empirically for the reconstruction of 3d human motion in monocular video sequences. Our tests on both real and motion capture-based sequences show significant performance gains with respect to competing nearest-neighbor, regression, and structured prediction methods.

Human motion and action analysis in video is an actively growing field with a broad spectrum of applications including video browsing and indexing, entertainment, virtual reality, human-computer interaction and surveillance. However, human motion analysis systems face important scientific and computational challenges. The proportions of the human body vary across individuals due to gender, weight, age or race. Aside from this variability, any single human body has many degrees of freedom due to articulation and the individual limbs are deformable due to muscle and clothing. Finally many realworld scenes involve multiple interacting humans occluded by each other or by other objects, and the scene conditions may also vary due to the camera motion or lighting changes. These factors make accurate 3D human models difficult to build and difficult to reconstruct reliably from 'flat' 2D images. During this talk I will discuss learning and inference algorithms for estimating 3D human motion in monocular video. While the problem has been traditionally approached using the powerful machinery of generative models, operating in an analysis by synthesis loop, the main emphasis of this talk will be on an emerging class of complementary discriminative temporal estimation models. These can be viewed as upside down, bottom-up versions of classical temporal models used with Kalman filtering or particle filtering. But instead of inverting a generative imaging model, we will learn to cooperatively predict complex, feedforward 2D-to-3D mappings, using Conditional Bayesian Mixtures of Experts. These are embedded in a probabilistic temporal framework in order to enforce dynamic constraints and allow a principled propagation of uncertainty. We call the resulting model BM3E (a Conditional Bayesian Mixture of Experts Markov Model). During the talk, I will discuss how inference can be restricted, for efficiency, to low-dimensional non-linear state spaces, and how the framework can be extended in order to deal with clutter and occlusion. I will also discuss the relative advantages of generative and initialize and recover from failure.

We present a new method for training deformable models. Assume that we have training images where part locations have been labeled. Typically, one fits a model by maximizing the likelihood of the part labels. Alternatively, one could fit a model such that, when the model is run on the training images, it finds the parts. We do this by maximizing the conditional likelihood of the training data. We formulate model-learning as parameter estimation in a conditional random field (CRF). Initializing parameters with their maximum likelihood estimates, we reach the global optimum by gradient ascent. We present a learning algorithm that searches exhaustively over all part locations in an image without relying on feature detectors. This provides millions of examples of training data, and seems to avoid over-fitting issues known with CRFs. Results for part localization are relatively scarce in the community. We present results on three established datasets; Caltech motorbikes [8], USC people [19], and Weizmann horses [3]. In the Caltech set we significantly outperform the state-of-the-art [6]. For the challenging people dataset, we present results that are comparable to [19], but are obtained using a significantly more generic model (devoid of a face or skin detector). Our model is general enough to find other articulated objects; we use it to recover poses of horses in the challenging Weizmann database.

We present an algorithm for jointly learning a consistent bidirectional generative-recognition model that combines top-down and bottom-up processing for monocular 3d human motion reconstruction. Learning progresses in alternative stages of self-training that optimize the probability of the image evidence: the recognition model is tunned using samples from the generative model and the generative model is optimized to produce inferences close to the ones predicted by the current recognition model. At equilibrium, the two models are consistent. During on-line inference, we scan the image at multiple locations and predict 3d human poses using the recognition model. But this implicitly includes one-shot generative consistency feedback. The framework provides a uniform treatment of human detection, 3d initialization and 3d recovery from transient failure. Our experimental results show that this procedure is promising for the automatic reconstruction of human motion in more natural scene settings with background clutter and occlusion.

We describe a mixture density propagation algorithm to estimate 3D human motion in monocular video sequences based on observations encoding the appearance of image silhouettes. Our approach is discriminative rather than generative, therefore it does not require the probabilistic inversion of a predictive observation model. Instead, it uses a large human motion capture data-base and a 3D computer graphics human model in order to synthesize training pairs of typical human configurations together with their realistically rendered 2D silhouettes. These are used to directly learn to predict the conditional state distributions required for 3D body pose tracking and thus avoid using the generative 3D model for inference (the learned discriminative predictors can also be used, complementary, as importance samplers in order to improve mixing or initialize generative inference algorithms). We aim for probabilistically motivated tracking algorithms and for models that can represent complex multivalued mappings common in inverse, uncertain perception inferences. Our paper has three contributions: (1) we establish the density propagation rules for discriminative inference in continuous, temporal chain models; (2) we propose flexible algorithms for learning multimodal state distributions based on compact, conditional Bayesian mixture of experts models; and (3) we demonstrate the algorithms empirically on real and motion capture-based test sequences and compare against nearest-neighbor and regression methods.

We present algorithms for recognizing human motion in monocular video sequences, based on discriminative conditional random field (CRF) and maximum entropy Markov models (MEMM). Existing approaches to this problem typically use generative (joint) structures like the hidden Markov model (HMM). Therefore they have to make simplifying, often unrealistic assumptions on the conditional independence of observations given the motion class labels and cannot accommodate overlapping features or long term contextual dependencies in the observation sequence. In contrast, conditional models like the CRFs seamlessly represent contextual dependencies, support efficient, exact inference using dynamic programming, and their parameters can be trained using convex optimization. We introduce conditional graphical models as complementary tools for human motion recognition and present an extensive set of experiments that show how these typically outperform HMMs in classifying not only diverse human activities like walking, jumping. running, picking or dancing, but also for discriminating among subtle motion styles like normal walk and wander walk

We present a conditional temporal probabilistic framework for recon- structing 3D human motion in monocular video based on descriptors en- coding image silhouette observations. For computational efficiency we restrict visual inference to low-dimensional kernel induced non-linear state spaces. Our methodology (kBME) combines kernel PCA-based non-linear dimensionality reduction (kPCA) and Conditional Bayesian Mixture of Experts (BME) in order to learn complex multivalued pre- dictors between observations and model hidden states. This is necessary for accurate, inverse, visual perception inferences, where several proba- ble, distant 3D solutions exist due to noise or the uncertainty of monoc- ular perspective projection. Low-dimensional models are appropriate because many visual processes exhibit strong non-linear correlations in both the image observations and the target, hidden state variables. The learned predictors are temporally combined within a conditional graphi- cal model in order to allow a principled propagation of uncertainty. We study several predictors and empirically show that the proposed algo- rithm positively compares with techniques based on regression, Kernel Dependency Estimation (KDE) or PCA alone, and gives results competi- tive to those of high-dimensional mixture predictors at a fraction of their computational cost. We show that the method successfully reconstructs the complex 3D motion of humans in real monocular video sequences. 1 Introduction and Related Work We consider the problem of inferring 3D articulated human motion from monocular video. This research topic has applications for scene understanding including human-computer in- terfaces, markerless human motion capture, entertainment and surveillance. A monocular approach is relevant because in real-world settings the human body parts are rarely com- pletely observed even when using multiple cameras. This is due to occlusions form other people or objects in the scene. A robust system has to necessarily deal with incomplete, ambiguous and uncertain measurements. Methods for 3D human motion reconstruction can be classified as generative and discriminative. They both require a state representation, namely a 3D human model with kinematics (joint angles) or shape (surfaces or joint po- sitions) and they both use a set of image features as observations for state inference. The computational goal in both cases is the conditional distribution for the model state given image observations. Generative model-based approaches [6, 16, 14, 13] have been demonstrated to flexibly re- construct complex unknown human motions and to naturally handle problem constraints. However it is difficult to construct reliable observation likelihoods due to the complexity of modeling human appearance. This varies widely due to different clothing and defor- mation, body proportions or lighting conditions. Besides being somewhat indirect, the generative approach further imposes strict conditional independence assumptions on the temporal observations given the states in order to ensure computational tractability. Due to these factors inference is expensive and produces highly multimodal state distributions [6, 16, 13]. Generative inference algorithms require complex annealing schedules [6, 13] or systematic non-linear search for local optima [16] in order to ensure continuing tracking. These difficulties motivate the advent of a complementary class of discriminative algo- rithms [10, 12, 18, 2], that approximate the state conditional directly, in order to simplify inference. However, inverse, observation-to-state multivalued mappings are difficult to learn (see e.g. fig. 1a) and a probabilistic temporal setting is necessary. In an earlier paper [15] we introduced a probabilistic discriminative framework for human motion reconstruc- tion. Because the method operates in the originally selected state and observation spaces that can be task generic, therefore redundant and often high-dimensional, inference is more expensive and can be less robust. To summarize, reconstructing 3D human motion in a Figure 1: (a, Left) Example of 180o ambiguity in predicting 3D human poses from sil- houette image features (center). It is essential that multiple plausible solutions (e.g. F1 and F2) are correctly represented and tracked over time. A single state predictor will either average the distant solutions or zig-zag between them, see also tables 1 and 2. (b, Right) A conditional chain model. The local distributions p(ytjyt(cid:0)1; zt) or p(ytjzt) are learned as in fig. 2. For inference, the predicted local state conditional is recursively combined with the filtered prior c.f . (1). conditional temporal framework poses the following difficulties: (i) The mapping between temporal observations and states is multivalued (i.e. the local conditional distributions to be learned are multimodal), therefore it cannot be accurately represented using global function approximations. (ii) Human models have multivariate, high-dimensional continuous states of 50 or more human joint angles. The temporal state conditionals are multimodal which makes efficient Kalman filtering algorithms inapplicable. General inference methods (par- ticle filters, mixtures) have to be used instead, but these are expensive for high-dimensional models (e.g. when reconstructing the motion of several people that operate in a joint state space). (iii) The components of the human state and of the silhouette observation vector ex- hibit strong correlations, because many repetitive human activities like walking or running have low intrinsic dimensionality. It appears wasteful to work with high-dimensional states of 50+ joint angles. Even if the space were truly high-dimensional, predicting correlated state dimensions independently may still be suboptimal. In this paper we present a conditional temporal estimation algorithm that restricts visual inference to low-dimensional, kernel induced state spaces. To exploit correlations among observations and among state variables, we model the local, temporal conditional distri- butions using ideas from Kernel PCA [11, 19] and conditional mixture modeling [7, 5], here adapted to produce multiple probabilistic predictions. The corresponding predictor is referred to as a Conditional Bayesian Mixture of Low-dimensional Kernel-Induced Experts (kBME). By integrating it within a conditional graphical model framework (fig. 1b), we can exploit temporal constraints probabilistically. We demonstrate that this methodology is effective for reconstructing the 3D motion of multiple people in monocular video. Our con- tribution w.r.t. [15] is a probabilistic conditional inference framework that operates over a non-linear, kernel-induced low-dimensional state spaces, and a set of experiments (on both real and artificial image sequences) that show how the proposed framework positively com- pares with powerful predictors based on KDE, PCA, or with the high-dimensional models of [15] at a fraction of their cost. 2 Probabilistic Inference in a Kernel Induced State Space We work with conditional graphical models with a chain structure [9], as shown in fig. 1b, These have continuous temporal states yt, t = 1 : : : T , observations zt. For compactness, we denote joint states Yt = (y1; y2; : : : ; yt) or joint observations Zt = (z1; : : : ; zt). Learning and inference are based on local conditionals: p(ytjzt) and p(ytjyt(cid:0)1; zt), with yt and zt being low-dimensional, kernel induced representations of some initial model having state xt and observation rt. We obtain zt; yt from rt, xt using kernel PCA [11, 19]. Inference is performed in a low-dimensional, non-linear, kernel induced latent state space (see fig. 1b and fig. 2 and (1)). For display or error reporting, we compute the original conditional p(xjr), or a temporally filtered version p(xtjRt); Rt = (r1; r2; : : : ; rt), using a learned pre-image state map [3]. 2.1 Density Propagation for Continuous Conditional Chains For online filtering, we compute the optimal distribution p(ytjZt) for the state yt, con- ditioned by observations Zt up to time t. The filtered density can be recursively derived as: p(ytjZt) = Zyt(cid:0)1 p(ytjyt(cid:0)1; zt)p(yt(cid:0)1jZt(cid:0)1) (1) We compute using a conditional mixture for p(ytjyt(cid:0)1; zt) (a Bayesian mixture of experts c.f . x2.2) and the prior p(yt(cid:0)1jZt(cid:0)1), each having, say M components. We integrate M 2 pairwise products of Gaussians analytically. The means of the expanded posterior are clus- tered and the centers are used to initialize a reduced M-component Kullback-Leibler ap- proximation that is refined using gradient descent [15]. The propagation rule (1) is similar to the one used for discrete state labels [9], but here we work with multivariate continuous state spaces and represent the local multimodal state conditionals using kBME (fig. 2), and not log-linear models [9] (these would require intractable normalization). This complex continuous model rules out inference based on Kalman filtering or dynamic programming [9]. 2.2 Learning Bayesian Mixtures over Kernel Induced State Spaces (kBME) In order to model conditional mappings between low-dimensional non-linear spaces we rely on kernel dimensionality reduction and conditional mixture predictors. The authors of KDE [19] propose a powerful structured unimodal predictor. This works by decorrelating the output using kernel PCA and learning a ridge regressor between the input and each decorrelated output dimension. Our procedure is also based on kernel PCA but takes into account the structure of the studied visual problem where both inputs and outputs are likely to be low-dimensional and the mapping between them multivalued. The output variables xi are projected onto the column vectors of the principal space in order to obtain their principal coordinates yi. A

We present a model-based framework for incremental, adaptive object shape estimation and tracking in monocular image sequences. Parametric structure and motion estimation methods usually assume a fixed class of shape representation (splines, deformable superquadrics, etc.) that is initialized prior to tracking. Since the model shape coverage is fixed a priori, the incremental recovery of structure is decoupled from tracking, thereby limiting both processes in their scope and robustness. In this work, we describe a model-based framework that supports the automatic detection and integration of low-level geometric primitives (lines) incrementally. Such primitives are not explicitly captured in the initial model, but are moving consistently with its image motion. The consistency tests used to identify new structure are based on trinocular constraints between geometric primitives. The method allows not only an increase in the model scope, but also improves tracking accuracy by including the newly recovered features in its state estimation. The formulation is a step toward automatic model building, since it allows both weaker assumptions on the availability of a prior shape representation and on the number of features that would otherwise be necessary for entirely bottom-up reconstruction. We demonstrate the proposed approach on two separate image-based tracking domains, each involving complex 3D object structure and motion.

Many difficult visual perception problems, like 3D human motion estimation, can be formulated in terms of inference using complex generative models, defined over high-dimensional state spaces. Despite progress, optimizing such models is difficult because prior knowledge cannot be flexibly integrated in order to reshape an initially designed representation space. Nonlinearities, inherent sparsity of high-dimensional training sets, and lack of global continuity makes dimensionality reduction challenging and low-dimensional search inefficient. To address these problems, we present a learning and inference algorithm that restricts visual tracking to automatically extracted, non-linearly embedded, low-dimensional spaces. This formulation produces a layered generative model with reduced state representation, that can be estimated using efficient continuous optimization methods. Our prior flattening method allows a simple analytic treatment of low-dimensional intrinsic curvature constraints, and allows consistent interpolation operations. We analyze reduced manifolds for human interaction activities, and demonstrate that the algorithm learns continuous generative models that are useful for tracking and for the reconstruction of 3D human motion in monocular video.

A major difficulty for 3D (three-dimensional) human body tracking from monocular image sequences is the near nonobservability of kinematic degrees of freedom that generate motion in depth. For known link (body segment) lengths, the strict nonobservabilities reduce to twofold 'forwards/backwards flipping' ambiguities for each link. These imply 2/sup # links/ formal inverse kinematics solutions for the full model, and hence linked groups of O(2/sup # links/) local minima in the model-image matching cost function. Choosing the wrong minimum leads to rapid mistracking, so for reliable tracking, rapid methods of investigating alternative minima within a group are needed. Previous approaches to this have used generic search methods that do not exploit the specific problem structure. Here, we complement these by using simple kinematic reasoning to enumerate the tree of possible forwards/backwards flips, thus greatly speeding the search within each linked group of minima. Our methods can be used either deterministically, or within stochastic 'jump-diffusion' style search processes. We give experimental results on some challenging monocular human tracking sequences, showing how the new kinematic-flipping based sampling method improves and complements existing ones.

Previous approaches to deformable model shape estimation and tracking have assumed a fixed class of shapes representation (e.g., deformable superquadrics), initialized prior to tracking. Since the shape coverage of the model is fixed, such approaches do not directly accommodate incremental representation discovery during tracking. As a result, model shape coverage is decoupled from tracking, thereby limiting both processes in terms of scope and robustness. We present a novel deformable model framework that accommodates the incremental incorporation during tracking of new geometric primitives (lines, in addition to points) that are not explicitly captured in the initial deformable model but that are moving consistently with its image motion. As these new features are detected via consistency checks, they are added to the model, providing incremental soft constraints on the estimation of its rigid parameters. The consistency checks are based on trilinear relationships between geometric primitives. Consequently, we not only increase both model scope and, ultimately, its higher-level shape coverage, but improve tracking robustness and accuracy, by directly employing the new features in both forward prediction and reconstruction. Our new formulation is a step towards automating model shape estimation and tracking, since it requires significantly reduced initial model hand-crafting. We demonstrate our approach on two separate image-based tracking domains, each involving complex 3D object shape and motion.

We present a method for recovering 3D human body motion from monocular video sequences using robust image matching, joint limits and non-self-intersection constraints, and a new sample-and-refine search strategy guided by rescaled cost-function covariances. Monocular 3D body tracking is challenging: for reliable tracking at least 30 joint parameters need to be estimated, subject to highly nonlinear physical constraints; the problem is chronically ill conditioned as about 1/3 of the d.o.f. (the depth-related ones) are almost unobservable in any given monocular image; and matching an imperfect, highly flexible self-occluding model to cluttered image features is intrinsically hard. To reduce correspondence ambiguities we use a carefully designed robust matching-cost metric that combines robust optical flow, edge energy, and motion boundaries. Even so, the ambiguity, nonlinearity and non-observability make the parameter-space cost surface multi-modal, unpredictable and ill conditioned, so minimizing it is difficult. We discuss the limitations of CONDENSATION-like samplers, and introduce a novel hybrid search algorithm that combines inflated-covariance-scaled sampling and continuous optimization subject to physical constraints. Experiments on some challenging monocular sequences show that robust cost modelling, joint and self-intersection constraints, and informed sampling are all essential for reliable monocular 3D body tracking.