We revisit the role of texture in monocular 3D hand reconstruction, not as an afterthought for photorealism, but as a dense, spatially grounded cue that can actively support pose and shape estimation. Our observation is simple: even in high-performing models, the overlay between predicted hand geometry and image appearance is often imperfect, suggesting that texture alignment may be an underused supervisory signal. We propose a lightweight texture module that embeds per-pixel observations into UV texture space and enables a novel dense alignment loss between predicted and observed hand appearances. Our approach assumes access to a differentiable rendering pipeline and a model that maps images to 3D hand meshes with known topology, allowing us to back-project a textured hand onto the image and perform pixel-based alignment. The module is self-contained and easily pluggable into existing reconstruction pipelines. To isolate and highlight the value of texture-guided supervision, we augment HaMeR [47], a high-performing yet unadorned transformer architecture for 3D hand pose estimation. The resulting system improves both accuracy and realism, demonstrating the value of appearance-guided alignment in hand reconstruction. Code, models and weights are available at https://gkarv.github.io/hand-texture-module

We present HandGAN (H-GAN), a cycle-consistent adversarial learning approach implementing multi-scale perceptual discriminators. It is designed to translate synthetic images of hands to the real domain. Synthetic hands provide complete ground-truth annotations, yet they are not representative of the target distribution of real-world data. We strive to provide the perfect blend of a realistic hand appearance with synthetic annotations. Relying on image-to-image translation, we improve the appearance of synthetic hands to approximate the statistical distribution underlying a collection of real images of hands. H-GAN tackles not only the cross-domain tone mapping but also structural differences in localized areas such as shading discontinuities. Results are evaluated on a qualitative and quantitative basis improving previous works. Furthermore, we relied on the hand classification task to claim our generated hands are statistically similar to the real domain of hands.

We present a clock-driven Spiking Neural Network simulator which is up to 3x faster than the state of the art while, at the same time, being more general and requiring less programming effort on both the user's and maintainer's side. This is made possible by designing our pipeline around "work queues" which act as interfaces between stages and greatly reduce implementation complexity. We evaluate our work using three well-established SNN models on a series of benchmarks.

This work demonstrates a real-time 3D hand tracking application that runs via computation offloading. The proposed framework enables the application to run on low-end mobile devices such as laptops and tablets, despite the fact that they lack the sufficient hardware to perform the required computations locally. The network connection takes the place of a GPGPU accelerator and sharing resources with a larger workstation becomes the acceleration mechanism. The unique properties of a generative optimizer are examined and constitute a challenging use-case, since the requirement for real-time performance makes it very latency-sensitive.

We consider the problem of estimating realistic contact forces during manipulation, backed with ground-truth measurements, using vision alone. Interaction forces are usually measured by mounting force transducers onto the manipulated objects or the hands. Those are costly, cumbersome, and alter the objects' physical properties and their perception by the human sense of touch. Our work establishes that interaction forces can be estimated in a cost-effective, reliable, non-intrusive way using vision. This is a complex and challenging problem. Indeed, in multi-contact, a given motion can generally be caused by an infinity of possible force distributions. To alleviate the limitations of traditional models based on inverse optimization, we collect and release the first large-scale dataset on manipulation kinodynamics as 3.2 hours of synchronized force and motion measurements under 193 object-grasp configurations. We learn a mapping between high-level kinematic features based on the equations of motion and the underlying manipulation forces using recurrent neural networks (RNN). The RNN predictions are consistently refined using physics-based optimization through second-order cone programming (SOCP). We show that our method can successfully capture interaction forces compatible with both the observations and the way humans intuitively manipulate objects, using a single RGB-D camera.

The analysis and the understanding of object manipulation scenarios based on computer vision techniques can be greatly facilitated if we can gain access to the full articulation of the manipulating hands and the 3D pose of the manipulated objects. Currently, there exist methods for tracking hands in interaction with objects whose 3D models are known [2]. There are also methods that can reconstruct 3D models of objects that are partially observable in each frame of a sequence [3]. However, no method can track hands in interaction with unknown objects, ie objects whose 3D model is not known a priori. In this paper we propose a novel approach that can track human hands in interaction with unknown objects. As illustrated in Fig.1, the input to the method is a sequence of RGBD frames showing the interaction of one or two hands with an unknown object. Starting with the raw depth map (left) we perform a pre-processing step and compute the scene point cloud. We employ an appropriately modified model based hand tracker [4] and temporal information to track the hand 3D positions and posture (middle bottom). In this process, a progressively built object model is also taken into account to cope with hand-object occlusions. We use the estimated fingertip positions of the hand to segment the manipulated object from the rest of the scene (middle top). The segmented object points are used to update the object position and orientation in the current frame and are integrated into the object 3D representation (right). More specifically, the work flow of the proposed approach consists of five main components linked together as shown in Fig. 2. At a first, preprocessing stage, the raw depth information from the sensor is prepared to enter the pipeline. A point cloud is computed along with the normals for each vertex. Then, the user’s hands are tracked in the scene. An articulated model for the left and right hands, with 26 degrees of freedom each, is fit to the pre-processed depth input. The current, possibly incomplete (or even empty, for the first frame) object model is incorporated to hand tracking to assist in handling hand/object occlusions. Using the computed 3D location of the user’s hands as well as the last position of the (possibly incomplete) object model, the region of the object is segmented in the input depth map. The hands are masked-out from the observation, by comparing it to the rendered hand models. Object tracking is achieved using a mutli-scale ICP [1]. The segmented object depth is used for a coarse to fine alignment with the (partially reconstructed) object model. Finally, the segmented and aligned depth data of the object with the current, partial 3D model are merged. The object’s 3D model is maintained in a voxel grid with a Truncated Signed Distance Function (TSDF) [3] representation. Experiment Proposed [2], GT model [2], Scanned model mean/median error mean/median error mean/median error Single hand, cat 0.42 / 0.39 0.47 / 0.43 0.45 / 0.43 Single hand, spray 0.65 / 0.63 0.70 / 0.53 0.63 / 0.47 Two hands, cat 0.38 / 0.34 0.33 / 0.31 0.44 / 0.39 Two hands, spray 0.59 / 0.44 0.51 / 0.38 0.62 / 0.41 Table 1: Hand tracking accuracy (in cm) measured on the synthetic datasets. The accuracy of the method is close to that of [2], although the latter assumes that the object model is known a priori.

3D tracking of objects and hands in an object manipulation scenario is a very interesting computer vision problem with a wide variety of applications ranging from consumer electronics to robotics and medicine. Recent advances in this research topic allow for 3D tracking of complex scenarios involving bi-manual manipulation of several rigid objects using commodity hardware and with high accuracy. The problem with these approaches is that they treat tracking as a search problem whose dimensionality increases with the number of objects in the scene. This fact typically limits the number of the tracked objects and/or the processing framerate. In this paper we present a method that utilizes simple low level motion cues for dynamically assigning computational resources to parts of the scene where they are actually required. In a series of experiments, we show that this simple idea improves tracking performance dramatically at a cost of only a minor degradation of tracking accuracy. The works that are most related to ours are the approaches by Kyriazis and Argyros on top-down 3D tracking of multiple active objects from RGBD input [1, 2]. The methodological part of our contribution can be briefly described as an extra processing node in the pipeline of [2] which it extends. The tracking approach in [2], the Ensemble of Collaborative Trackers (ECT), regards a set of semi-independent trackers. Each tracker is associated with a distinct object in the scene. For an object to be tracked, a separate optimization problem is solved, one for each frame and each object. Each optimization problem is numerically solved using a black box optimizer, i.e. a variant of the Particle Swarm Optimization (PSO) algorithm. PSO treats the objective function as an oracle and queries it on purposefully evolved “guesses” in the search space, in order to find the optimum. The harder the problem is, the more guesses (budget, which is the product of PSO particles and PSO generations that need to be computed) are required for adequate accuracy to be achieved. For example, tracking the pose and the articulation of the hand amounts to solving a 27-parameter optimization problem. This is much harder than solving for the 3D pose of a rigid object (6 parameters). ECT [2] assigns a fixed amount of computational resources to each tracking sub-problem which depends on this notion of complexity and which is empirically estimated. In this work, we quantify at run time, how hard the tracking of an object should be, not only based on its intrinsic complexity, but also based on its observed dynamics. An object that appears static in the recent temporal window requires less resources compared to an object whose state is more dynamic. Thus, change detection is performed on the image space of color intensities and depth measurements of the RGBD input. In more detail, for the next tracking frame, and the ith tracker, a value mi is computed, which takes the value of 0 if the corresponding tracked object appears to be relatively static, and takes the value of 1 otherwise. For mi to be 1, any of the following need to be true: • the mean value of the pixel-wise differences between the observations of object i and the back-projection of the last estimated configuration of object i is high enough, • the kinetic energy of object i, as computed from so far tracked velocities, suggests a moving object, • the amount of missing depth measurements changes substantially, from one frame to the next, which might me attributed to change in the slant of object i, or • any of the above was satisfied in the recent past (damping). Budget has the trivial minimum value of 0. Moreover, as it has been shown in [3, 4], a PSO budget of 64 particles and 64 generations suffices to track a hand, even in interaction with another hand. No more budget should be required to track simpler structures such as rigid objects. The proposed dynamic budget allocation policy assigns a minimum budget Bmin of 64 particles running for 4 generations to the objects that are static and a maximum budget Bmax that never exceeds the aforementioned (a) Pouring pancake mix (b) Waiting for the mix to cook

We consider the problem of tracking multiple interacting objects in 3D, using RGBD input and by considering a hypothesize-and-test approach. Due to their interaction, objects to be tracked are expected to occlude each other in the field of view of the camera observing them. A naive approach would be to employ a Set of Independent Trackers (SIT) and to assign one tracker to each object. This approach scales well with the number of objects but fails as occlusions become stronger due to their disjoint consideration. The solution representing the current state of the art employs a single Joint Tracker (JT) that accounts for all objects simultaneously. This directly resolves ambiguities due to occlusions but has a computational complexity that grows geometrically with the number of tracked objects. We propose a middle ground, namely an Ensemble of Collaborative Trackers (ECT), that combines best traits from both worlds to deliver a practical and accurate solution to the multi-object 3D tracking problem. We present quantitative and qualitative experiments with several synthetic and real world sequences of diverse complexity. Experiments demonstrate that ECT manages to track far more complex scenes than JT at a computational time that is only slightly larger than that of SIT.

In several hand-object(s) interaction scenarios, the change in the objects' state is a direct consequence of the hand's motion. This has a straightforward representation in Newtonian dynamics. We present the first approach that exploits this observation to perform model-based 3D tracking of a table-top scene comprising passive objects and an active hand. Our forward modelling of 3D hand-object(s) interaction regards both the appearance and the physical state of the scene and is parameterized over the hand motion (26 DoFs) between two successive instants in time. We demonstrate that our approach manages to track the 3D pose of all objects and the 3D pose and articulation of the hand by only searching for the parameters of the hand motion. In the proposed framework, covert scene state is inferred by connecting it to the overt state, through the incorporation of physics. Thus, our tracking approach treats a variety of challenging observability issues in a principled manner, without the need to resort to heuristics.

In this paper we use a Hierarchical Hidden Markov Model (HHMM) to represent and learn complex activities/task performed by humans/robots in everyday life. Action primitives are used as a grammar to represent complex human behaviour and learn the interactions and behaviour of human/robots with different objects. The main contribution is the use of a probabilistic model capable of representing behaviours at multiple levels of abstraction to support the proposed hypothesis. The hierarchical nature of the model allows decomposition of the complex task into simple action primitives. The framework is evaluated with data collected for tasks of everyday importance performed by a human user.

The main contribution of this paper is a probabilistic method for predicting human manipulation intention from image sequences of human-object interaction. Predicting intention amounts to inferring the imminent manipulation task when human hand is observed to have stably grasped the object. Inference is performed by means of a probabilistic graphical model that encodes object grasping tasks over the 3D state of the observed scene. The 3D state is extracted from RGB-D image sequences by a novel vision-based, markerless hand-object 3D tracking framework. To deal with the high-dimensional state-space and mixed data types (discrete and continuous) involved in grasping tasks, we introduce a generative vector quantization method using mixture models and self-organizing maps. This yields a compact model for encoding of grasping actions, able of handling uncertain and partial sensory data. Experimentation showed that the model trained on simulated data can provide a potent basis for accurate goal-inference with partial and noisy observations of actual real-world demonstrations. We also show a grasp selection process, guided by the inferred human intention, to illustrate the use of the system for goal-directed grasp imitation.

We propose a method that relies on markerless visual observations to track the full articulation of two hands that interact with each-other in a complex, unconstrained manner. We formulate this as an optimization problem whose 54-dimensional parameter space represents all possible configurations of two hands, each represented as a kinematic structure with 26 Degrees of Freedom (DoFs). To solve this problem, we employ Particle Swarm Optimization (PSO), an evolutionary, stochastic optimization method with the objective of finding the two-hands configuration that best explains observations provided by an RGB-D sensor. To the best of our knowledge, the proposed method is the first to attempt and achieve the articulated motion tracking of two strongly interacting hands. Extensive quantitative and qualitative experiments with simulated and real world image sequences demonstrate that an accurate and efficient solution of this problem is indeed feasible.

We present a novel solution to the problem of recovering and tracking the 3D position, orientation and full articulation of a human hand from markerless visual observations obtained by a Kinect sensor. We treat this as an optimization problem, seeking for the hand model parameters that minimize the discrepancy between the appearance and 3D structure of hypothesized instances of a hand model and actual hand observations. This optimization problem is effectively solved using a variant of Particle Swarm Optimization (PSO). The proposed method does not require special markers and/or a complex image acquisition setup. Being model based, it provides continuous solutions to the problem of tracking hand articulations. Extensive experiments with a prototype GPU-based implementation of the proposed method demonstrate that accurate and robust 3D tracking of hand articulations can be achieved in near real-time (15Hz).

Due to occlusions, the estimation of the full pose of a human hand interacting with an object is much more challenging than pose recovery of a hand observed in isolation. In this work we formulate an optimization problem whose solution is the 26-DOF hand pose together with the pose and model parameters of the manipulated object. Optimization seeks for the joint hand-object model that (a) best explains the incompleteness of observations resulting from occlusions due to hand-object interaction and (b) is physically plausible in the sense that the hand does not share the same physical space with the object. The proposed method is the first that solves efficiently the continuous, full-DOF, joint hand-object tracking problem based solely on markerless multicamera input. Additionally, it is the first to demonstrate how hand-object interaction can be exploited as a context that facilitates hand pose estimation, instead of being considered as a complicating factor. Extensive quantitative and qualitative experiments with simulated and real world image sequences as well as a comparative evaluation with a state-of-the-art method for pose estimation of isolated hands, support the above findings.