Large Language Models (LLMs) have shown promise in the autonomous driving sector, particularly in generalization and interpretability. We introduce a unique objectlevel multimodal LLM architecture that merges vectorized numeric modalities with a pre-trained LLM to improve context understanding in driving situations. We also present a new dataset of 160k QA pairs derived from 10k driving scenarios, paired with high quality control commands collected with RL agent and question answer pairs generated by teacher LLM (GPT-3.5). A distinct pretraining strategy is devised to align numeric vector modalities with static LLM representations using vector captioning language data. We also introduce an evaluation metric for Driving QA and demonstrate our LLM-driver’s proficiency in interpreting driving scenarios, answering questions, and decision-making. Our findings highlight the potential of LLM-based driving action generation in comparison to traditional behavioral cloning. We make our benchmark, datasets, and model available1for further exploration.

An accurate model of the environment and the dynamic agents acting in it offers great potential for improving motion planning. We present MILE: a Model-based Imitation LEarning approach to jointly learn a model of the world and a policy for autonomous driving. Our method leverages 3D geometry as an inductive bias and learns a highly compact latent space directly from high-resolution videos of expert demonstrations. Our model is trained on an offline corpus of urban driving data, without any online interaction with the environment. MILE improves upon prior state-of-the-art by 31% in driving score on the CARLA simulator when deployed in a completely new town and new weather conditions. Our model can predict diverse and plausible states and actions, that can be interpretably decoded to bird's-eye view semantic segmentation. Further, we demonstrate that it can execute complex driving manoeuvres from plans entirely predicted in imagination. Our approach is the first camera-only method that models static scene, dynamic scene, and ego-behaviour in an urban driving environment. The code and model weights are available at https://github.com/wayveai/mile.

The increased availability and maturity of head-mounted and wearable devices opens up opportunities for remote communication and collaboration. However, the signal streams provided by these devices (e.g., head pose, hand pose, and gaze direction) do not represent a whole person. One of the main open problems is therefore how to leverage these signals to build faithful representations of the user. In this paper, we propose a method based on variational autoencoders to generate articulated poses of a human skeleton based on noisy streams of head and hand pose. Our approach relies on a model of pose likelihood that is novel and theoretically well-grounded. We demonstrate on publicly available datasets that our method is effective even from very impoverished signals and investigate how pose prediction can be made more accurate and realistic.

Recent work on Neural Radiance Fields (NeRF) showed how neural networks can be used to encode complex 3D environments that can be rendered photorealistically from novel viewpoints. Rendering these images is very computationally demanding and recent improvements are still a long way from enabling interactive rates, even on high-end hardware. Motivated by scenarios on mobile and mixed reality devices, we propose FastNeRF, the first NeRF-based system capable of rendering high fidelity photorealistic images at 200Hz on a high-end consumer GPU. The core of our method is a graphics-inspired factorization that allows for (i) compactly caching a deep radiance map at each position in space, (ii) efficiently querying that map using ray directions to estimate the pixel values in the rendered image. Extensive experiments show that the proposed method is 3000 times faster than the original NeRF algorithm and at least an order of magnitude faster than existing work on accelerating NeRF, while maintaining visual quality and extensibility.

We demonstrate that it is possible to perform face-related computer vision in the wild using synthetic data alone. The community has long enjoyed the benefits of synthesizing training data with graphics, but the domain gap between real and synthetic data has remained a problem, especially for human faces. Researchers have tried to bridge this gap with data mixing, domain adaptation, and domain-adversarial training, but we show that it is possible to synthesize data with minimal domain gap, so that models trained on synthetic data generalize to real in-the-wild datasets. We describe how to combine a procedurally-generated parametric 3D face model with a comprehensive library of hand-crafted assets to render training images with unprecedented realism and diversity. We train machine learning systems for face-related tasks such as landmark localization and face parsing, showing that synthetic data can both match real data in accuracy as well as open up new approaches where manual labeling would be impossible.

In this work we present an articulated tracking approach for robotic manipulators, which relies only on visual cues from colour and depth images to estimate the robot's state when interacting with or being occluded by its environment. We hypothesise that articulated model fitting approaches can only achieve accurate tracking if subpixel-level accurate correspondences between observed and estimated state can be established. Previous work in this area has exclusively relied on either discriminative depth information or colour edge correspondences as tracking objective and required initialisation from joint encoders. In this paper we propose a coarse-to-fine articulated state estimator, which relies only on visual cues from colour edges and learned depth keypoints, and which is initialised from a robot state distribution predicted from a depth image. We evaluate our approach on four RGB-D sequences showing a KUICA LWR arm with a Schunk SDH2 hand interacting with its environment and demonstrate that this combined keypoint and edge tracking objective can estimate the palm position with an average error of 2. 5cm without using any joint encoder sensing.

Hand pose estimation, formulated as an inverse problem, is typically optimized by an energy function over pose parameters using a 'black box' image generation procedure, knowing little about either the relationships between the parameters or the form of the energy function. In this paper, we show significant improvement upon such black box optimization by exploiting high-level knowledge of the parameter structure and using a local surrogate energy function. Our new framework, hierarchical sampling optimization (HSO), consists of a sequence of discriminative predictors organized into a kinematic hierarchy. Each predictor is conditioned on its ancestors, and generates a set of samples over a subset of the pose parameters, with only one selected by the highly-efficient surrogate energy. The selected partial poses are concatenated to generate a full-pose hypothesis. Repeating the same process, several hypotheses are generated and the full energy function selects the best result. Under the same kinematic hierarchy, two methods based on decision forest and convolutional neural network are proposed to generate the samples and two optimization methods are studied when optimizing these samples. Experimental evaluations on three publicly available datasets show that our method is particularly impressive in low-compute scenarios where it significantly outperforms all other state-of-the-art methods.

This paper focuses on visual tracking of a robotic manipulator during manipulation. In this situation, tracking is prone to failure when visual distractions are created by the object being manipulated and the clutter in the environment. Current state-of-the-art approaches, which typically rely on model-fitting using Iterative Closest Point (ICP), fail in the presence of distracting data points and are unable to recover. Meanwhile, discriminative methods which are trained only to distinguish parts of the tracked object can also fail in these scenarios as data points from the occlusions are incorrectly classified as being from the manipulator. We instead propose to use the per-pixel data-to-model associations provided from a random forest to avoid local minima during model fitting. By training the random forest with artificial occlusions we can achieve increased robustness to occlusion and clutter present in the scene. We do this without specific knowledge about the type or location of the manipulated object. Our approach is demonstrated by using dense depth data from an RGB-D camera to track a robotic manipulator during manipulation and in presence of occlusions.

RANSAC is an important algorithm in robust optimization and a central building block for many computer vision applications. In recent years, traditionally hand-crafted pipelines have been replaced by deep learning pipelines, which can be trained in an end-to-end fashion. However, RANSAC has so far not been used as part of such deep learning pipelines, because its hypothesis selection procedure is non-differentiable. In this work, we present two different ways to overcome this limitation. The most promising approach is inspired by reinforcement learning, namely to replace the deterministic hypothesis selection by a probabilistic selection for which we can derive the expected loss w.r.t. to all learnable parameters. We call this approach DSAC, the differentiable counterpart of RANSAC. We apply DSAC to the problem of camera localization, where deep learning has so far failed to improve on traditional approaches. We demonstrate that by directly minimizing the expected loss of the output camera poses, robustly estimated by RANSAC, we achieve an increase in accuracy. In the future, any deep learning pipeline can use DSAC as a robust optimization component.

State-of-the-art computer vision algorithms often achieve efficiency by making discrete choices about which hypotheses to explore next. This allows allocation of computational resources to promising candidates, however, such decisions are non-differentiable. As a result, these algorithms are hard to train in an end-to-end fashion. In this work we propose to learn an efficient algorithm for the task of 6D object pose estimation. Our system optimizes the parameters of an existing state-of-the art pose estimation system using reinforcement learning, where the pose estimation system now becomes the stochastic policy, parametrized by a CNN. Additionally, we present an efficient training algorithm that dramatically reduces computation time. We show empirically that our learned pose estimation procedure makes better use of limited resources and improves upon the state-of-the-art on a challenging dataset. Our approach enables differentiable end-to-end training of complex algorithmic pipelines and learns to make optimal use of a given computational budget.

We present a fast, practical method for personalizing a hand shape basis to an individual user's detailed hand shape using only a small set of depth images. To achieve this, we minimize an energy based on a sum of render-and-compare cost functions called the golden energy. However, this energy is only piecewise continuous, due to pixels crossing occlusion boundaries, and is therefore not obviously amenable to efficient gradient-based optimization. A key insight is that the energy is the combination of a smooth low-frequency function with a high-frequency, low-amplitude, piecewisecontinuous function. A central finite difference approximation with a suitable step size can therefore jump over the discontinuities to obtain a good approximation to the energy's low-frequency behavior, allowing efficient gradient-based optimization. Experimental results quantitatively demonstrate for the first time that detailed personalized models improve the accuracy of hand tracking and achieve competitive results in both tracking and model registration.

This paper presents ShadowHands - a novel technique for visualizing a remote user's hand gestures using a single depth sensor and hand tracking system. Previous work has shown that making distributed users better aware of each other's gestures facilitates remote collaboration. These systems presented virtual embodiments as a stream of raw 2D or 3D data -- this data is noisy, and requires high bandwidth and favorable camera positions. Instead, our work uses a hand tracker to capture gestures which we visualize with a high-fidelity hand model. Our system is practical, requiring only a single depth sensor placed below the screen, and can be used without per-user calibration. As we use a 3D model rather than raw data, we can augment the hand's appearance to improve saliency and aesthetics. We alpha-blend this visualization over a shared workspace, so the local user perceives the remote user's hand as if they were separated by a transparent display. We conducted an experiment to compare traditional hand embodiments with our new technique, showing a quantitative improvement in selection accuracy and qualitative improvements in feelings of mutual understanding and engagement.

The sixteen papers in this special section focus on human pose recovery and behavior analysis (HuPBA). This is one of the most challenging topics in computer vision, pattern analysis, and machine learning. It is of critical importance for application areas that include gaming, computer interaction, human robot interaction, security, commerce, assistive technologies and rehabilitation, sports, sign language recognition, and driver assistance technology, to mention just a few. In essence, HuPBA requires dealing with the articulated nature of the human body, changes in appearance due to clothing, and the inherent problems of clutter scenes, such as background artifacts, occlusions, and illumination changes. These papers represent the most recent research in this field, including new methods considering still images, image sequences, depth data, stereo vision, 3D vision, audio, and IMUs, among others.

Fully articulated hand tracking promises to enable fundamentally new interactions with virtual and augmented worlds, but the limited accuracy and efficiency of current systems has prevented widespread adoption. Today's dominant paradigm uses machine learning for initialization and recovery followed by iterative model-fitting optimization to achieve a detailed pose fit. We follow this paradigm, but make several changes to the model-fitting, namely using: (1) a more discriminative objective function; (2) a smooth-surface model that provides gradients for non-linear optimization; and (3) joint optimization over both the model pose and the correspondences between observed data points and the model surface. While each of these changes may actually increase the cost per fitting iteration, we find a compensating decrease in the number of iterations. Further, the wide basin of convergence means that fewer starting points are needed for successful model fitting. Our system runs in real-time on CPU only, which frees up the commonly over-burdened GPU for experience designers. The hand tracker is efficient enough to run on low-power devices such as tablets. We can track up to several meters from the camera to provide a large working volume for interaction, even using the noisy data from current-generation depth cameras. Quantitative assessments on standard datasets show that the new approach exceeds the state of the art in accuracy. Qualitative results take the form of live recordings of a range of interactive experiences enabled by this new approach.

We present a new real-time hand tracking system based on a single depth camera. The system can accurately reconstruct complex hand poses across a variety of subjects. It also allows for robust tracking, rapidly recovering from any temporary failures. Most uniquely, our tracker is highly flexible, dramatically improving upon previous approaches which have focused on front-facing close-range scenarios. This flexibility opens up new possibilities for human-computer interaction with examples including tracking at distances from tens of centimeters through to several meters (for controlling the TV at a distance), supporting tracking using a moving depth camera (for mobile scenarios), and arbitrary camera placements (for VR headsets). These features are achieved through a new pipeline that combines a multi-layered discriminative reinitialization strategy for per-frame pose estimation, followed by a generative model-fitting stage. We provide extensive technical details and a detailed qualitative and quantitative analysis.

We describe how to learn a compact and efficient model of the surface deformation of human hands. The model is built from a set of noisy depth images of a diverse set of subjects performing different poses with their hands. We represent the observed surface using Loop subdivision of a control mesh that is deformed by our learned parametric shape and pose model. The model simultaneously accounts for variation in subject-specific shape and subject-agnostic pose. Specifically, hand shape is parameterized as a linear combination of a mean mesh in a neutral pose with a small number of offset vectors. This mesh is then articulated using standard linear blend skinning (LBS) to generate the control mesh of a subdivision surface. We define an energy that encourages each depth pixel to be explained by our model, and the use of a smooth subdivision surface allows us to optimize for all parameters jointly from a rough initialization. The efficacy of our method is demonstrated using both synthetic and real data, where it is shown that hand shape variation can be represented using only a small number of basis components. We compare with other approaches including PCA and show a substantial improvement in the representational power of our model, while maintaining the efficiency of a linear shape basis.

Recent advances in camera relocalization use predictions from a regression forest to guide the camera pose optimization procedure. In these methods, each tree associates one pixel with a point in the scene's 3D world coordinate frame. In previous work, these predictions were point estimates and the subsequent camera pose optimization implicitly assumed an isotropic distribution of these estimates. In this paper, we train a regression forest to predict mixtures of anisotropic 3D Gaussians and show how the predicted uncertainties can be taken into account for continuous pose optimization. Experiments show that our proposed method is able to relocalize up to 40% more frames than the state of the art.

Consumer depth cameras have dramatically improved our ability to track rigid, articulated, and deformable 3D objects in real-time. However, depth cameras have a limited temporal resolution (frame-rate) that restricts the accuracy and robustness of tracking, especially for fast or unpredictable motion. In this paper, we show how to perform model-based object tracking which allows to reconstruct the object's depth at an order of magnitude higher frame-rate through simple modifications to an off-the-shelf depth camera. We focus on phase-based time-of-flight (ToF) sensing, which reconstructs each low frame-rate depth image from a set of short exposure 'raw' infrared captures. These raw captures are taken in quick succession near the beginning of each depth frame, and differ in the modulation of their active illumination. We make two contributions. First, we detail how to perform model-based tracking against these raw captures. Second, we show that by reprogramming the camera to space the raw captures uniformly in time, we obtain a 10x higher frame-rate, and thereby improve the ability to track fast-moving objects.

Hand pose estimation has matured rapidly in recent years. The introduction of commodity depth sensors and a multitude of practical applications have spurred new advances. We provide an extensive analysis of the state-of-the-art, focusing on hand pose estimation from a single depth frame. To do so, we have implemented a considerable number of systems, and will release all software and evaluation code. We summarize important conclusions here: (1) Pose estimation appears roughly solved for scenes with isolated hands. However, methods still struggle to analyze cluttered scenes where hands may be interacting with nearby objects and surfaces. To spur further progress we introduce a challenging new dataset with diverse, cluttered scenes. (2) Many methods evaluate themselves with disparate criteria, making comparisons difficult. We define a consistent evaluation criteria, rigorously motivated by human experiments. (3) We introduce a simple nearest-neighbor baseline that outperforms most existing systems. This implies that most systems do not generalize beyond their training sets. This also reinforces the under-appreciated point that training data is as important as the model itself. We conclude with directions for future progress.

We address the problem of hand pose estimation, formulated as an inverse problem. Typical approaches optimize an energy function over pose parameters using a 'black box' image generation procedure. This procedure knows little about either the relationships between the parameters or the form of the energy function. In this paper, we show that we can significantly improving upon black box optimization by exploiting high-level knowledge of the structure of the parameters and using a local surrogate energy function. Our new framework, hierarchical sampling optimization, consists of a sequence of predictors organized into a kinematic hierarchy. Each predictor is conditioned on its ancestors, and generates a set of samples over a subset of the pose parameters. The highly-efficient surrogate energy is used to select among samples. Having evaluated the full hierarchy, the partial pose samples are concatenated to generate a full-pose hypothesis. Several hypotheses are generated using the same procedure, and finally the original full energy function selects the best result. Experimental evaluation on three publically available datasets show that our method is particularly impressive in low-compute scenarios where it significantly outperforms all other state-of-the-art methods.

In this paper we report the set-up and results of the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) organized in conjunction with the MICCAI 2012 and 2013 conferences. Twenty state-of-the-art tumor segmentation algorithms were applied to a set of 65 multi-contrast MR scans of low- and high-grade glioma patients-manually annotated by up to four raters-and to 65 comparable scans generated using tumor image simulation software. Quantitative evaluations revealed considerable disagreement between the human raters in segmenting various tumor sub-regions (Dice scores in the range 74%-85%), illustrating the difficulty of this task. We found that different algorithms worked best for different sub-regions (reaching performance comparable to human inter-rater variability), but that no single algorithm ranked in the top for all sub-regions simultaneously. Fusing several good algorithms using a hierarchical majority vote yielded segmentations that consistently ranked above all individual algorithms, indicating remaining opportunities for further methodological improvements. The BRATS image data and manual annotations continue to be publicly available through an online evaluation system as an ongoing benchmarking resource.

We present a new interactive and online approach to 3D scene understanding. Our system, SemanticPaint , allows users to simultaneously scan their environment whilst interactively segmenting the scene simply by reaching out and touching any desired object or surface. Our system continuously learns from these segmentations, and labels new unseen parts of the environment. Unlike offline systems where capture, labeling, and batch learning often take hours or even days to perform, our approach is fully online. This provides users with continuous live feedback of the recognition during capture, allowing to immediately correct errors in the segmentation and/or learning—a feature that has so far been unavailable to batch and offline methods. This leads to models that are tailored or personalized specifically to the user's environments and object classes of interest, opening up the potential for new applications in augmented reality, interior design, and human/robot navigation. It also provides the ability to capture substantial labeled 3D datasets for training large-scale visual recognition systems.

Recovery from tracking failure is essential in any simultaneous localization and tracking system. In this context, we explore an efficient keyframe-based relocalization method based on frame encoding using randomized ferns. The method enables automatic discovery of keyframes through online harvesting in tracking mode, and fast retrieval of pose candidates in the case when tracking is lost. Frame encoding is achieved by applying simple binary feature tests which are stored in the nodes of an ensemble of randomized ferns. The concatenation of small block codes generated by each fern yields a global compact representation of camera frames. Based on those representations we define the frame dissimilarity as the block-wise hamming distance (BlockHD). Dissimilarities between an incoming query frame and a large set of keyframes can be efficiently evaluated by simply traversing the nodes of the ferns and counting image co-occurrences in corresponding code tables. In tracking mode, those dissimilarities decide whether a frame/pose pair is considered as a novel keyframe. For tracking recovery, poses of the most similar keyframes are retrieved and used for reinitialization of the tracking algorithm. The integration of our relocalization method into a hand-held KinectFusion system allows seamless continuation of mapping even when tracking is frequently lost.

We present RetroDepth, a new vision-based system for accurately sensing the 3D silhouettes of hands, styluses, and other objects, as they interact on and above physical surfaces. Our setup is simple, cheap, and easily reproducible, comprising of two infrared cameras, diffuse infrared LEDs, and any off-the-shelf retro-reflective material. The retro-reflector aids image segmentation, creating a strong contrast between the surface and any object in proximity. A new highly efficient stereo matching algorithm precisely estimates the 3D contours of interacting objects and the retro-reflective surfaces. A novel pipeline enables 3D finger, hand and object tracking, as well as gesture recognition, purely using these 3D contours. We demonstrate high-precision sensing, allowing robust disambiguation between a finger or stylus touching, pressing or interacting above the surface. This allows many interactive scenarios that seamlessly mix together freehand 3D interactions with touch, pressure and stylus input. As shown, these rich modalities of input are enabled on and above any retro-reflective surface, including custom "physical widgets" fabricated by users. We compare our system with Kinect and Leap Motion, and conclude with limitations and future work.

We propose 'filter forests' (FF), an efficient new discriminative approach for predicting continuous variables given a signal and its context. FF can be used for general signal restoration tasks that can be tackled via convolutional filtering, where it attempts to learn the optimal filtering kernels to be applied to each data point. The model can learn both the size of the kernel and its values, conditioned on the observation and its spatial or temporal context. We show that FF compares favorably to both Markov random field based and recently proposed regression forest based approaches for labeling problems in terms of efficiency and accuracy. In particular, we demonstrate how FF can be used to learn optimal denoising filters for natural images as well as for other tasks such as depth image refinement, and 1D signal magnitude estimation. Numerous experiments and quantitative comparisons show that FFs achieve accuracy at par or superior to recent state of the art techniques, while being several orders of magnitude faster.

We address the problem of estimating the pose of a cam- era relative to a known 3D scene from a single RGB-D frame. We formulate this problem as inversion of the generative rendering procedure, i.e., we want to find the camera pose corresponding to a rendering of the 3D scene model that is most similar with the observed input. This is a non-convex optimization problem with many local optima. We propose a hybrid discriminative-generative learning architecture that consists of: (i) a set of M predictors which generate M camera pose hypotheses, and (ii) a 'selector' or 'aggregator' that infers the best pose from the multiple pose hypotheses based on a similarity function. We are interested in predictors that not only produce good hypotheses but also hypotheses that are different from each other. Thus, we propose and study methods for learning 'marginally relevant' predictors, and compare their performance when used with different selection procedures. We evaluate our method on a recently released 3D reconstruction dataset with challenging camera poses, and scene variability. Experiments show that our method learns to make multiple predictions that are marginally relevant and can effectively select an accurate prediction. Furthermore, our method outperforms the state-of-the-art discriminative approach for camera relocalization.

This paper presents a method for acquiring dense nonrigid shape and deformation from a single monocular depth sensor. We focus on modeling the human hand, and assume that a single rough template model is available. We combine and extend existing work on model-based tracking, subdivision surface fitting, and mesh deformation to acquire detailed hand models from as few as 15 frames of depth data. We propose an objective that measures the error of fit between each sampled data point and a continuous model surface defined by a rigged control mesh, and uses as-rigid-as-possible (ARAP) regularizers to cleanly separate the model and template geometries. A key contribution is our use of a smooth model based on subdivision surfaces that allows simultaneous optimization over both correspondences and model parameters. This avoids the use of iterated closest point (ICP) algorithms which often lead to slow convergence. Automatic initialization is obtained using a regression forest trained to infer approximate correspondences. Experiments show that the resulting meshes model the user's hand shape more accurately than just adapting the shape parameters of the skeleton, and that the retargeted skeleton accurately models the user's articulations. We investigate the effect of various modeling choices, and show the benefits of using subdivision surfaces and ARAP regularization.

We present a machine learning technique for estimating absolute, per-pixel depth using any conventional monocular 2D camera, with minor hardware modifications. Our approach targets close-range human capture and interaction where dense 3D estimation of hands and faces is desired. We use hybrid classification-regression forests to learn how to map from near infrared intensity images to absolute , metric depth in real-time. We demonstrate a variety of human-computer interaction and capture scenarios. Experiments show an accuracy that outperforms a conventional light fall-off baseline, and is comparable to high-quality consumer depth cameras, but with a dramatically reduced cost, power consumption, and form-factor.

Traditionally, human pose estimation algorithms could be classified into generative [2] and discriminative [4] approaches. Generative approaches model the likelihood of the observations given a pose estimate, however, they are susceptible to local minima and thus require good initial pose estimates. Discriminative approaches learn a direct mapping from image features to pose space from training data, however, they struggle to generalize to unseen poses. Building on previous work [3], Taylor et al. [5] bypass some of these limitations using a hybrid-approach that discriminatively predicts, for each pixel in a depth image, a corresponding point on the surface of a humanoid mesh model. This mesh model is then robustly fit to the resulting set of correspondences using local optimization. Surprisingly though, these correspondences are actually inferred using a random forest whose structure was trained using a classification objective that arbitrarily equates target model points belonging to the same predefined body part [3]. In this paper, we address Taylor et al.’s use of this proxy classification objective by proposing Metric Space Information Gain (MSIG), a replacement objective function for training a random forest to directly minimize the uncertainty over the target model points, naturally encoding the correlation between these points as a function of the geodesic distance. To this end, we view the surface of the model U as a metric space (U,dU) defined by the geodesic distance metric dU (see first panel of Figure 1). The natural objective function to minimize the uncertainty in the resulting true distributions that result from a split function s in such a space, is the information gain I(s) [1]. This is generally approximated using an empirical distribution Q = {ui} ⊆U drawn from the true unsplit distribution pU as I(s)≈ I(s;Q) = Ĥ(Q)− ∑ i∈{L,R} |Qi| |Q| Ĥ(Qi), (1)

Conventional decision forest based methods for image labelling tasks like object segmentation make predictions for each variable (pixel) independently [3, 5, 8]. This prevents them from enforcing dependencies between variables and translates into locally inconsistent pixel labellings. Random field models, instead, encourage spatial consistency of labels at increased computational expense. This paper presents a new and efficient forest based model that achieves spatially consistent semantic image segmentation by encoding variable dependencies directly in the feature space the forests operate on. Such correlations are captured via new long-range, soft connectivity features, computed via generalized geodesic distance transforms. Our model can be thought of as a generalization of the successful Semantic Texton Forest, Auto-Context, and Entangled Forest models. A second contribution is to show the connection between the typical Conditional Random Field (CRF) energy and the forest training objective. This analysis yields a new objective for training decision forests that encourages more accurate structured prediction. Our GeoF model is validated quantitatively on the task of semantic image segmentation, on four challenging and very diverse image datasets. GeoF outperforms both state of-the-art forest models and the conventional pair wise CRF.

We address the problem of inferring the pose of an RGB-D camera relative to a known 3D scene, given only a single acquired image. Our approach employs a regression forest that is capable of inferring an estimate of each pixel's correspondence to 3D points in the scene's world coordinate frame. The forest uses only simple depth and RGB pixel comparison features, and does not require the computation of feature descriptors. The forest is trained to be capable of predicting correspondences at any pixel, so no interest point detectors are required. The camera pose is inferred using a robust optimization scheme. This starts with an initial set of hypothesized camera poses, constructed by applying the forest at a small fraction of image pixels. Preemptive RANSAC then iterates sampling more pixels at which to evaluate the forest, counting inliers, and refining the hypothesized poses. We evaluate on several varied scenes captured with an RGB-D camera and observe that the proposed technique achieves highly accurate relocalization and substantially out-performs two state of the art baselines.

Image segmentation is the process of partitioning an image into segments or subsets of pixels for purposes of further analysis, such as separating the interesting objects in the foreground from the un-interesting objects in the background. In many image processing applications, the process requires a sequence of computational steps on a per pixel basis, thereby binding the performance to the size and resolution of the image. As applications require greater resolution and larger images the computational resources of this step can quickly exceed those of available CPUs, especially in the power and thermal constrained areas of consumer electronics and mobile. In this work, we use a hardware tree-based classifier to solve the image segmentation problem. The application is background removal (BGR) from depth-maps obtained from the Microsoft Kinect sensor. After the image is segmented, subsequent steps then classify the objects in the scene. The approach is flexible: to address different application domains we only need to change the trees used by the classifiers. We describe two distinct approaches and evaluate their performance using the commercial-grade testing environment used for the Microsoft Xbox gaming console.

We introduce new approaches for augmenting annotated training datasets used for object detection tasks that serve achieving two goals: reduce the effort needed for collecting and manually annotating huge datasets and introduce novel variations to the initial dataset that help the learning algorithms. The methods presented in this work aim at relocating objects using their segmentation masks to new backgrounds. These variations comprise changes in properties of objects such as spatial location in the image, surrounding context and scale. We propose a model selection approach to arbitrate between the constructed model on a per class basis. Experimental results show gains that can be harvested using the proposed approach.

We introduce an efficient camera relocalization approach which can be easily integrated into real-time 3D reconstruction methods, such as KinectFusion. Our approach makes use of compact encoding of whole image frames which enables both online harvesting of keyframes in tracking mode, and fast retrieval of pose proposals when tracking is lost. The encoding scheme is based on randomized ferns and simple binary feature tests. Each fern generates a small block code, and the concatenation of codes yields a compact representation of each camera frame. Based on those representations we introduce an efficient frame dissimilarity measure which is defined via the block-wise hamming distance (BlockHD). We illustrate how BlockHDs between a query frame and a large set of keyframes can be simultaneously evaluated by traversing the nodes of the ferns and counting image co-occurrences in corresponding code tables. In tracking mode, this mechanism allows us to consider every frame/pose pair as a potential keyframe. A new keyframe is added only if it is sufficiently dissimilar from all previously stored keyframes. For tracking recovery, camera poses are retrieved that correspond to the keyframes with smallest BlockHDs. The pose proposals are then used to reinitialize the tracking algorithm. Harvesting of keyframes and pose retrieval are computationally efficient with only small impact on the run-time performance of the 3D reconstruction. Integrating our relocalization method into KinectFusion allows seamless continuation of mapping even when tracking is frequently lost. Additionally, we demonstrate how marker-free augmented reality, in particular, can benefit from this integration by enabling a smoother and continuous AR experience.

Randomized decision trees and forests have a rich history in machine learning and have seen considerable success in application, perhaps particularly so for computer vision. However, they face a fundamental limitation: given enough data, the number of nodes in decision trees will grow exponentially with depth. For certain applications, for example on mobile or embedded processors, memory is a limited resource, and so the exponential growth of trees limits their depth, and thus their potential accuracy. This paper proposes decision jungles, revisiting the idea of ensembles of rooted decision directed acyclic graphs (DAGs), and shows these to be compact and powerful discriminative models for classification. Unlike conventional decision trees that only allow one path to every node, a DAG in a decision jungle allows multiple paths from the root to each leaf. We present and compare two new node merging algorithms that jointly optimize both the features and the structure of the DAGs efficiently. During training, node splitting and node merging are driven by the minimization of exactly the same objective function, here the weighted sum of entropies at the leaves. Results on varied datasets show that, compared to decision forests and several other baselines, decision jungles require dramatically less memory while considerably improving generalization.

We describe two new approaches to human pose estimation. Both can quickly and accurately predict the 3D positions of body joints from a single depth image without using any temporal information. The key to both approaches is the use of a large, realistic, and highly varied synthetic set of training images. This allows us to learn models that are largely invariant to factors such as pose, body shape, field-of-view cropping, and clothing. Our first approach employs an intermediate body parts representation, designed so that an accurate per-pixel classification of the parts will localize the joints of the body. The second approach instead directly regresses the positions of body joints. By using simple depth pixel comparison features and parallelizable decision forests, both approaches can run super-real time on consumer hardware. Our evaluation investigates many aspects of our methods, and compares the approaches to each other and to the state of the art. Results on silhouettes suggest broader applicability to other imaging modalities.

With the invention of the low-cost Microsoft Kinect sensor, high-resolution depth and visual (RGB) sensing has become available for widespread use. The complementary nature of the depth and visual information provided by the Kinect sensor opens up new opportunities to solve fundamental problems in computer vision. This paper presents a comprehensive review of recent Kinect-based computer vision algorithms and applications. The reviewed approaches are classified according to the type of vision problems that can be addressed or enhanced by means of the Kinect sensor. The covered topics include preprocessing, object tracking and recognition, human activity analysis, hand gesture analysis, and indoor 3-D mapping. For each category of methods, we outline their main algorithmic contributions and summarize their advantages/differences compared to their RGB counterparts. Finally, we give an overview of the challenges in this field and future research trends. This paper is expected to serve as a tutorial and source of references for Kinect-based computer vision researchers.

Kinect sensor, high-resolution depth and visual (RGB) sensing has become available for widespread use as an off-the-shelf technology. This special issue is specifically dedicated to new algorithms and/or new applications based on the Kinect (or similar RGB-D) sensors. In total, we received over ninety submissions from more than twenty countries all around the world. The submissions cover a wide range of areas including object and scene classification, 3-D pose estimation, visual tracking, data fusion, human action/activity recognition, 3-D reconstruction, mobile robotics, and so on. After two rounds of review by at least two (mostly three) expert reviewers for each paper, the Guest Editors have selected twelve high-quality papers to be included in this highly popular special issue. The papers that comprise this issue are briefly summarized.

Random forests have been successfully applied to various high level computer vision tasks such as human pose estimation and object segmentation. These models are extremely efficient but work under the assumption that the output variables (such as body part locations or pixel labels) are independent. In this paper, we present a conditional regression forest model for human pose estimation that incorporates dependency relationships between output variables through a global latent variable while still maintaining a low computational cost. We show that the incorporation of a global latent variable encoding torso orientation, or human height, etc., can dramatically increase the accuracy of body joint location prediction. Our model also allows efficient and seamless incorporation of prior knowledge about the problem instance such as the height or orientation of the human subject which can be available from the problem context or via a temporal model. We show that our method significantly outperforms state-of-the-art methods for pose estimation from depth images. The conditional regression model proposed in the paper is general and can be applied to other problems where random forests are used.

Fitting an articulated model to image data is often approached as an optimization over both model pose and model-to-image correspondence. For complex models such as humans, previous work has required a good initialization, or an alternating minimization between correspondence and pose. In this paper we investigate one-shot pose estimation: can we directly infer correspondences using a regression function trained to be invariant to body size and shape, and then optimize the model pose just once? We evaluate on several challenging single-frame data sets containing a wide variety of body poses, shapes, torso rotations, and image cropping. Our experiments demonstrate that one-shot pose estimation achieves state of the art results and runs in real-time.

We propose a new method to quickly and accurately predict 3D positions of body joints from a single depth image, using no temporal information. We take an object recognition approach, designing an intermediate body parts representation that maps the difficult pose estimation problem into a simpler per-pixel classification problem. Our large and highly varied training dataset allows the classifier to estimate body parts invariant to pose, body shape, clothing, etc. Finally we generate confidence-scored 3D proposals of several body joints by reprojecting the classification result and finding local modes. The system runs at 200 frames per second on consumer hardware. Our evaluation shows high accuracy on both synthetic and real test sets, and investigates the effect of several training parameters. We achieve state of the art accuracy in our comparison with related work and demonstrate improved generalization over exact whole-skeleton nearest neighbor matching.

We present a new approach to general-activity human pose estimation from depth images, building on Hough forests. We extend existing techniques in several ways: real time prediction of multiple 3D joints, explicit learning of voting weights, vote compression to allow larger training sets, and a comparison of several decision-tree training objectives. Key aspects of our work include: regression directly from the raw depth image, without the use of an arbitrary intermediate representation; applicability to general motions (not constrained to particular activities) and the ability to localize occluded as well as visible body joints. Experimental results demonstrate that our method produces state of the art results on several data sets including the challenging MSRC-5000 pose estimation test set, at a speed of about 200 frames per second. Results on silhouettes suggest broader applicability to other imaging modalities.

We present a system for accurate real-time mapping of complex and arbitrary indoor scenes in variable lighting conditions, using only a moving low-cost depth camera and commodity graphics hardware. We fuse all of the depth data streamed from a Kinect sensor into a single global implicit surface model of the observed scene in real-time. The current sensor pose is simultaneously obtained by tracking the live depth frame relative to the global model using a coarse-to-fine iterative closest point (ICP) algorithm, which uses all of the observed depth data available. We demonstrate the advantages of tracking against the growing full surface model compared with frame-to-frame tracking, obtaining tracking and mapping results in constant time within room sized scenes with limited drift and high accuracy. We also show both qualitative and quantitative results relating to various aspects of our tracking and mapping system. Modelling of natural scenes, in real-time with only commodity sensor and GPU hardware, promises an exciting step forward in augmented reality (AR), in particular, it allows dense surfaces to be reconstructed in real-time, with a level of detail and robustness beyond any solution yet presented using passive computer vision.

KinectFusion enables a user holding and moving a standard Kinect camera to rapidly create detailed 3D reconstructions of an indoor scene. Only the depth data from Kinect is used to track the 3D pose of the sensor and reconstruct, geometrically precise, 3D models of the physical scene in real-time. The capabilities of KinectFusion, as well as the novel GPU-based pipeline are described in full. Uses of the core system for low-cost handheld scanning, and geometry-aware augmented reality and physics-based interactions are shown. Novel extensions to the core GPU pipeline demonstrate object segmentation and user interaction directly in front of the sensor, without degrading camera tracking or reconstruction. These extensions are used to enable real-time multi-touch interactions anywhere, allowing any planar or non-planar reconstructed physical surface to be appropriated for touch.

Computer vision has grown tremendously in the past two decades. Despite all efforts, existing attempts at matching parts of the human visual system's extraordinary ability to understand visual scenes lack either scope or power. By combining the advantages of general low-level generative models and powerful layer-based and hierarchical models, this work aims at being a first step toward richer, more flexible models of images. After comparing various types of restricted Boltzmann machines (RBMs) able to model continuous-valued data, we introduce our basic model, the masked RBM, which explicitly models occlusion boundaries in image patches by factoring the appearance of any patch region from its shape. We then propose a generative model of larger images using a field of such RBMs. Finally, we discuss how masked RBMs could be stacked to form a deep model able to generate more complicated structures and suitable for various tasks such as segmentation or object recognition.

This paper proposes a novel method for recognizing faces degraded by blur using deblurring of facial images. The main issue is how to infer a Point Spread Function (PSF) representing the process of blur on faces. Inferring a PSF from a single facial image is an ill-posed problem. Our method uses learned prior information derived from a training set of blurred faces to make the problem more tractable. We construct a feature space such that blurred faces degraded by the same PSF are similar to one another. We learn statistical models that represent prior knowledge of predefined PSF sets in this feature space. A query image of unknown blur is compared with each model and the closest one is selected for PSF inference. The query image is deblurred using the PSF corresponding to that model and is thus ready for recognition. Experiments on a large face database (FERET) artificially degraded by focus or motion blur show that our method substantially improves the recognition performance compared to existing methods. We also demonstrate improved performance on real blurred images on the FRGC 1.0 face database. Furthermore, we show and explain how combining the proposed facial deblur inference with the local phase quantization (LPQ) method can further enhance the performance.

This paper proposes a novel method for deblurring facial images to recognize faces degraded by blur. The main problem is how to infer a point spread function (PSF) representing the process of blur. Inferring a PSF from a single facial image is an ill-posed problem. To make this problem more tractable, our method uses learned prior information derived from a training set of blurred facial images of several individuals. We construct a feature space such that blurred faces degraded by the same PSF are similar to one another and form a cluster. During training, we compute a statistical model of each PSF cluster in this feature space. For PSF inference we compare a query image of unknown blur with each model and select the closest one. Using the PSF corresponding to that model, the query image is deblurred, ready for recognition. Experiments on a standard face database artificially degraded by focus or motion blur show that our method substantially improves the recognition performance compared with state-of-the-art methods. We also demonstrate improved performance on real blurred images.

This paper proposes an efficient fusion of contour and texture cues for im-age categorization and object detection. Our work confirms and strengthens recent results that combining complementary feature types improves perfor-mance. We obtain a similar improvement in accuracy and additionally an improvement in efficiency. We use a boosting algorithm to learn models that use contour and texture features. Our main contributions are (i) the use of dense generic texture features to complement contour fragments, and (ii) a simple feature selection mechanism that includes the computational costs of features in order to learn a run-time efficient model. Our evaluation on 17 challenging and varied object classes confirms that the synergy of the two feature types performs significantly better than either alone, and that computational efficiency is substantially improved using our feature selection mechanism. An investigation of the boosted features shows a fascinating emergent property: the absence of certain textures often con-tributes towards object detection. Comparison with recent work shows that performance is state of the art. 1

We propose semantic texton forests, efficient and powerful new low-level features. These are ensembles of decision trees that act directly on image pixels, and therefore do not need the expensive computation of filter-bank responses or local descriptors. They are extremely fast to both train and test, especially compared with k-means clustering and nearest-neighbor assignment of feature descriptors. The nodes in the trees provide (i) an implicit hierarchical clustering into semantic textons, and (ii) an explicit local classification estimate. Our second contribution, the bag of semantic textons, combines a histogram of semantic textons over an image region with a region prior category distribution. The bag of semantic textons is computed over the whole image for categorization, and over local rectangular regions for segmentation. Including both histogram and region prior allows our segmentation algorithm to exploit both textural and semantic context. Our third contribution is an image-level prior for segmentation that emphasizes those categories that the automatic categorization believes to be present. We evaluate on two datasets including the very challenging VOC 2007 segmentation dataset. Our results significantly advance the state-of-the-art in segmentation accuracy, and furthermore, our use of efficient decision forests gives at least a five-fold increase in execution speed.

Psychophysical studies [9], [17] show that we can recognize objects using fragments of outline contour alone. This paper proposes a new automatic visual recognition system based only on local contour features, capable of localizing objects in space and scale. The system first builds a class-specific codebook of local fragments of contour using a novel formulation of chamfer matching. These local fragments allow recognition that is robust to within-class variation, pose changes, and articulation. Boosting combines these fragments into a cascaded sliding-window classifier, and mean shift is used to select strong responses as a final set of detections. We show how learning can be performed iteratively on both training and test sets to boot-strap an improved classifier. We compare with other methods based on contour and local descriptors in our detailed evaluation over 17 challenging categories, and obtain highly competitive results. The results confirm that contour is indeed a powerful cue for multi-scale and multi-class visual object recognition.

This paper addresses the problem of detecting and segmenting partially occluded objects of a known category. We first define a part labelling which densely covers the object. Our Layout Consistent Random Field (LayoutCRF) model then imposes asymmetric local spatial constraints on these labels to ensure the consistent layout of parts whilst allowing for object deformation. Arbitrary occlusions of the object are handled by avoiding the assumption that the whole object is visible. The resulting system is both efficient to train and to apply to novel images, due to a novel annealed layout-consistent expansion move algorithm paired with a randomised decision tree classifier. We apply our technique to images of cars and faces and demonstrate state-of-the-art detection and segmentation performance even in the presence of partial occlusion.

Abstract Composite images are synthesized from existing photographs by artists who make concept art, e.g., storyboards for movies or architectural planning. Current techniques allow an artist to fabricate such an image by digitally splicing parts of stock photographs. While these images serve mainly to “quickly”convey how a scene should look, their production is laborious. We propose a technique that allows a person to design a new photograph with substantially less effort. This paper presents a method that generates a composite image when a user types in nouns, such as “boat”and “sand.”The artist can optionally design an intended image by specifying other constraints. Our algorithm formulates the constraints as queries to search an automatically annotated image database. The desired photograph, not a collage, is then synthesized using graph‐cut optimization, optionally allowing for further user interaction to edit or choose among alternative generated photos. An implementation of our approach, shown in the associated video, demonstrates our contributions of (1) a method for creating specific images with minimal human effort, and (2) a combined algorithm for automatically building an image library with semantic annotations from any photo collection.

We present a novel categorical object detection scheme that uses only local contour-based features. A two-stage, partially supervised learning architecture is proposed: a rudimentary detector is learned from a very small set of segmented images and applied to a larger training set of un-segmented images; the second stage bootstraps these detections to learn an improved classifier while explicitly training against clutter. The detectors are learned with a boosting algorithm which creates a location-sensitive classifier using a discriminative set of features from a randomly chosen dictionary of contour fragments. We present results that are very competitive with other state-of-the-art object detection schemes and show robustness to object articulations, clutter, and occlusion. Our major contributions are the application of boosted local contour-based features for object detection in a partially supervised learning framework, and an efficient new boosting procedure for simultaneously selecting features and estimating per-feature parameters.

A new algorithm is proposed for novel view generation in one-to-one teleconferencing applications. Given the video streams acquired by two cameras placed on either side of a computer monitor, the proposed algorithm synthesizes images from a virtual camera in arbitrary position (typically located within the monitor) to facilitate eye contact. Our technique is based on an improved, dynamic-programming, stereo algorithm for efficient novel-view generation. The two main contributions are: i) a new type of three-plane graph for dense-stereo dynamic-programming, that encourages correct occlusion labeling; ii) a compact geometric derivation for novel-view synthesis by direct projection of the minimum-cost surface. Furthermore, we present a novel algorithm for the temporal maintenance of a background model to enhance the rendering of occlusions and reduce temporal artefacts (flicker); and a cost aggregation algorithm that acts directly on our three-dimensional matching cost space. Examples are given that demonstrate the robustness of the new algorithm to spatial and temporal artefacts for long stereo video streams. These include demonstrations of synthesis of cyclopean views of extended conversational sequences. We further demonstrate synthesis from a freely translating virtual camera.