We introduce HyperGS, a novel framework for Hyper-Spectral Novel View Synthesis (HNVS), based on a new latent 3D Gaussian Splatting (3DGS) technique. Our approach enables simultaneous spatial and spectral renderings by encoding material properties from multi-view 3D hyperspectral datasets. HyperGS reconstructs high-fidelity views from arbitrary perspectives with improved accuracy and speed, outperforming currently existing methods. To address the challenges of high-dimensional data, we perform view synthesis in a learned latent space, incorporating a pixel-wise adaptive density function and a pruning technique for increased training stability and efficiency. Additionally, we introduce the first HNVS benchmark, implementing a number of new baselines based on recent SOTA RGB-NVS techniques, alongside the small number of prior works on HNVS. We demonstrate HyperGS's robustness through extensive evaluation of real and simulated hyper- spectral scenes with a 14db accuracy improvement upon previously published models.

We explore symbolic policy optimization for various legged locomotion challenges; specifically walker environments ranging from bipedal to highly redundant systems with 128 legs. These represent a broad range of action space dimensionalities. We find that state-of-the-art symbolic policy optimization approaches struggle to scale to these higher dimensional problems, due to the need to iterate over action dimensions, and their reliance on a neural network anchor policy. We thus propose Fast Symbolic Policy (FSP) to accelerate the training of symbolic locomotion policies. This approach avoids the need to iterate over the action dimensions, and does not require a pre-trained neural network anchor. We also propose Dim-X, a method for effectively reducing the action space dimensionality using the inductive priors of legged locomotion. We demonstrate that FSP with Dim-X can learn symbolic policies, with improved scaling performance compared to the baselines, vastly exceeding that possible with previous symbolic techniques. We further show that Dim-X on its own can also be integrated into neural network policies to shorten their training time and improve scaling performance.

One of the largest challenges in the deployment of legged robots in the real world is deriving effective general gaits. In this paper, we present BeeTLe, which is a framework that enables terrain aware locomotion without the need for dedicated terrain sensors. BeeTLe is realised as a multi-expert policy Reinforcement Learning (RL) algorithm. This enables multiple gaits, applicable to different surface types, to be stored and shared in a single policy. Sensor free terrain awareness is incorporated using a Recurrent Neural Network (RNN) to infer surface type purely from actuator positions over time. The RNN achieves an accuracy of 94% in terrain identification out of 8 possible options. We demonstrate that BeeTLe achieves a greater performance than the baselines across a series of challenges including: the traversal of a flat plane, a tilted plane, a sequence of tilted planes and geometry modelling a natural hilly terrain. This is despite not seeing the sequence of tilted planes and the natural hilly terrain during training.The code, policy and simulated environments are available at: https://gitlab.surrey.ac.uk/rf00350/BeeTLe

Cross-View Geo-Localisation within urban regions is challenging in part due to the lack of geospatial structuring within current datasets and techniques. We propose utilising graph representations to model sequences of local observations and the connectivity of the target location. Modelling as a graph enables generating previously unseen sequences by sampling with new parameter configurations. To leverage this newly available information, we propose a GNN-based architecture, producing spatially strong embeddings and improving discriminability over isolated image embeddings. We outline SpaG-BOL, introducing three novel contributions. 1) The first graph-structured dataset for Cross-View Geo-Localisation, containing multiple streetview images per node to improve generalisation. 2) Introducing GNNs to the problem, we develop the first system that exploits the correlation between node proximity and feature similarity. 3) Lever-aging the unique properties of the graph representation - we demonstrate a novel retrieval filtering approach based on neighbourhood bearings. SpaGBOL achieves state-of-the-art accuracies on the unseen test graph - with relative Top-1 retrieval improvements on previous techniques of 11%, and 50% when filtering with Bearing Vector Matching on the SpaGBOL dataset. Code and dataset available: github.com/tavisshore/SpaGBOL.

Cross-view image matching for geo-localisation is a challenging problem due to the significant visual difference between aerial and ground-level viewpoints. The method provides localisation capabilities from geo-referenced images, eliminating the need for external devices or costly equipment. This enhances the capacity of agents to autonomously determine their position, navigate, and operate effectively in GNSS-denied environments. Current research employs a variety of techniques to reduce the domain gap such as applying polar transforms to aerial images or synthesising between perspectives. However, these approaches generally rely on having a 360° field of view, limiting real-world feasibility. We propose BEV-CV, an approach introducing two key novelties with a focus on improving the real-world viability of cross-view geo-localisation. Firstly bringing ground-level images into a semantic Birds-Eye-View before matching embeddings, allowing for direct comparison with aerial image representations. Secondly, we adapt datasets into application realistic format - limited Field-of-View images aligned to vehicle direction. BEV-CV achieves state-of-the-art recall accuracies, improving Top-1 rates of 70° crops of CVUSA and CVACT by 23% and 24% respectively. Also decreasing computational requirements by reducing floating point operations to below previous works, and decreasing embedding dimensionality by 33% - together allowing for faster localisation capabilities.

Self-supervised monocular depth estimation (SS-MDE) has the potential to scale to vast quantities of data. Unfortunately, existing approaches limit themselves to the automotive domain, resulting in models incapable of generalizing to complex environments such as natural or indoor settings.To address this, we propose a large-scale SlowTV dataset curated from YouTube, containing an order of magnitude more data than existing automotive datasets. SlowTV contains 1.7M images from a rich diversity of environments, such as worldwide seasonal hiking, scenic driving and scuba diving. Using this dataset, we train an SS-MDE model that provides zero-shot generalization to a large collection of indoor/outdoor datasets. The resulting model outperforms all existing SSL approaches and closes the gap on supervised SoTA, despite using a more efficient architecture.We additionally introduce a collection of best-practices to further maximize performance and zero-shot generalization. This includes 1) aspect ratio augmentation, 2) camera intrinsic estimation, 3) support frame randomization and 4) flexible motion estimation. Code is available at https://github.com/jspenmar/slowtv_monodepth.

This paper presents a new information source for supporting robot localisation: material composition. The proposed method complements the existing visual, structural, and semantic cues utilized in the literature. However, it has a distinct advantage in its ability to differentiate structurally [23], visually [25] or categorically [1] similar objects such as different doors, by using Raman spectrometers. Such devices can identify the material of objects it probes through the bonds between the material's molecules. Unlike similar sensors, such as mass spectroscopy, it does so without damaging the material or environment. In addition to introducing the first material-based localisation algorithm, this paper supports the future growth of the field by presenting a gazebo plugin for Raman spectrometers, material sensing demonstrations, as well as the first-ever localisation data-set with benchmarks for material-based localisation. This benchmarking shows that the proposed technique results in a significant improvement over current state-of-the-art localisation techniques, achieving 16 % more accurate localisation than the leading baseline. The code and dataset will be released at: https://github.com/ThirgoodC/RaSpectLoc

In this work, we introduce a new perspective for learning transferable content in multi-task imitation learning. Humans are capable of transferring skills and knowledge. If we can cycle to work and drive to the store, we can also cycle to the store and drive to work. We take inspiration from this and hypothesize the latent memory of a policy network can be disentangled into two partitions. These contain either the knowledge of the environmental context for the task or the generalisable skill needed to solve the task. This allows an improved training efficiency and better generalization over previously unseen combinations of skills in the same environment, and the same task in unseen environments. We used the proposed approach to train a disentangled agent for two different multi-task IL environments. In both cases, we out-performed the SOTA by 30% in task success rate. We also demonstrated this for navigation on a real robot.

Robots need to be able to work in multiple different environments. Even when performing similar tasks, different behaviour should be deployed to best fit the current environment. In this paper, We propose a new approach to navigation, where it is treated as a multi-task learning problem. This enables the robot to learn to behave differently in visual navigation tasks for different environments while also learning shared expertise across environments. We evaluated our approach in both simulated environments as well as real-world data. Our method allows our system to converge with a 26% reduction in training time, while also increasing accuracy.

The use of reinforcement learning (RL) has led to huge advancements in the field of robotics. However data scarcity, brittle convergence and the gap between simulation & real world environments, mean that most common RL approaches are subject to over fitting and fail to generalise to unseen environments. Hardware agnostic policies would mitigate this by allowing a single network to operate in a variety of test domains, where dynamics vary due to changes in robotic morphologies or internal parameters. We utilise the idea that learning to adapt a known and successful control policy is easier and more flexible than jointly learning numerous control policies for different morphologies.This paper presents the idea of Hardware Agnostic Reinforcement Learning using Adversarial selection (HARL-A). In this approach training examples are sampled using a novel adversarial loss function. This is designed to self regulate morphologies based on their learning potential. Simply applying our learning potential based loss function to current state-of-the-art already provides ~ 30% improvement in performance. Meanwhile experiments using the full implementation of HARL-A report an average increase of 70% to a standard RL baseline and 55% compared with current state-of-the-art.

The successful performance of any system is dependant on the hardware of the agent, which is typically immutable during RL training. In this work, we present ORCHID (Optimisation of Robotic Control and Hardware In Design) which allows for truly simultaneous optimisation of hardware and control parameters in an RL pipeline. We show that by forming a complex differential path through a trajectory rollout we can leverage a vast amount of information from the system that was previously lost in the ‘black-box’ environment. Combining this with a novel hardware-conditioned critic network minimises variance during training and ensures stable updates are made. This allows for refinements to be made to both the morphology and control parameters simultaneously. The result is an efficient and versatile approach to holistic robot design, that brings the final system nearer to true optimality. We show improvements in performance across 4 different test environments with two different control algorithms - in all experiments the maximum performance achieved with ORCHID is shown to be unattainable using only policy updates with the default design. We also show how re-designing a robot using ORCHID in simulation, transfers to a vast improvement in the performance of a real-world robot.

The broad scope of obstacle avoidance has led to many kinds of computer vision-based approaches. Despite its popularity, it is not a solved problem. Traditional computer vision techniques using cameras and depth sensors often focus on static scenes, or rely on priors about the obstacles. Recent developments in bio-inspired sensors present event cameras as a compelling choice for dynamic scenes. Although these sensors have many advantages over their frame-based counterparts, such as high dynamic range and temporal resolution, event-based perception has largely remained in 2D. This often leads to solutions reliant on heuristics and specific to a particular task.We show that the fusion of events and depth overcomes the failure cases of each individual modality when performing obstacle avoidance. Our proposed approach unifies event camera and lidar streams to estimate metric Time-To-Impact (TTI) without prior knowledge of the scene geometry or obstacles. In addition, we release an extensive event-based dataset with six visual streams spanning over 700 scanned scenes.

Prior work on Sign Language Translation has shown that having a mid-level sign gloss representation (effectively recognizing the individual signs) improves the translation performance drastically. In fact, the current state-of-the-art in translation requires gloss level tokenization in order to work. We introduce a novel transformer based architecture that jointly learns Continuous Sign Language Recognition and Translation while being trainable in an end-to-end manner. This is achieved by using a Connectionist Temporal Classification (CTC) loss to bind the recognition and translation problems into a single unified architecture. This joint approach does not require any ground-truth timing information, simultaneously solving two co-dependant sequence-to-sequence learning problems and leads to significant performance gains. We evaluate the recognition and translation performances of our approaches on the challenging RWTH-PHOENIX-Weather-2014T (PHOENIX14T) dataset. We report state-of-the-art sign language recognition and translation results achieved by our Sign Language Transformers. Our translation networks outperform both sign video to spoken language and gloss to spoken language translation models, in some cases more than doubling the performance (9.58 vs. 21.80 BLEU-4 Score). We also share new baseline translation results using transformer networks for several other text-to-text sign language translation tasks.

"Like night and day" is a commonly used expression to imply that two things are completely different. Unfortunately, this tends to be the case for current visual feature representations of the same scene across varying seasons or times of day. The aim of this paper is to provide a dense feature representation that can be used to perform localization, sparse matching or image retrieval, regardless of the current seasonal or temporal appearance. Recently, there have been several proposed methodologies for deep learning dense feature representations. These methods make use of ground truth pixel-wise correspondences between pairs of images and focus on the spatial properties of the features. As such, they don't address temporal or seasonal variation. Furthermore, obtaining the required pixel-wise correspondence data to train in cross-seasonal environments is highly complex in most scenarios. We propose Deja-Vu, a weakly supervised approach to learning season invariant features that does not require pixel-wise ground truth data. The proposed system only requires coarse labels indicating if two images correspond to the same location or not. From these labels, the network is trained to produce "similar" dense feature maps for corresponding locations despite environmental changes. Code will be made available at: https://github.com/jspenmar/DejaVu_Features.

In the current monocular depth research, the dominant approach is to employ unsupervised training on large datasets, driven by warped photometric consistency. Such approaches lack robustness and are unable to generalize to challenging domains such as nighttime scenes or adverse weather conditions where assumptions about photometric consistency break down. We propose DeFeat-Net (Depth & Feature network), an approach to simultaneously learn a cross-domain dense feature representation, alongside a robust depth-estimation framework based on warped feature consistency. The resulting feature representation is learned in an unsupervised manner with no explicit ground-truth correspondences required. We show that within a single domain, our technique is comparable to both the current state of the art in monocular depth estimation and supervised feature representation learning. However, by simultaneously learning features, depth and motion, our technique is able to generalize to challenging domains, allowing DeFeat-Net to outperform the current state-of-the-art with around 10% reduction in all error measures on more challenging sequences such as nighttime driving.

We present a novel form of explanation for Reinforcement Learning, based around the notion of intended outcome. These explanations describe the outcome an agent is trying to achieve by its actions. We provide a simple proof that general methods for post-hoc explanations of this nature are impossible in traditional reinforcement learning. Rather, the information needed for the explanations must be collected in conjunction with training the agent. We derive approaches designed to extract local explanations based on intention for several variants of Q-function approximation and prove consistency between the explanations and the Q-values learned. We demonstrate our method on multiple reinforcement learning problems, and provide code to help researchers introspecting their RL environments and algorithms.

Abstract The use of human-level semantic information to aid robotic tasks has recently become an important area for both Computer Vision and Robotics. This has been enabled by advances in Deep Learning that allow consistent and robust semantic understanding. Leveraging this semantic vision of the world has allowed human-level understanding to naturally emerge from many different approaches. Particularly, the use of semantic information to aid in localisation and reconstruction has been at the forefront of both fields. Like robots, humans also require the ability to localise within a structure. To aid this, humans have designed high-level semantic maps of our structures called floorplans. We are extremely good at localising in them, even with limited access to the depth information used by robots. This is because we focus on the distribution of semantic elements, rather than geometric ones. Evidence of this is that humans are normally able to localise in a floorplan that has not been scaled properly. In order to grant this ability to robots, it is necessary to use localisation approaches that leverage the same semantic information humans use. In this paper, we present a novel method for semantically enabled global localisation. Our approach relies on the semantic labels present in the floorplan. Deep Learning is leveraged to extract semantic labels from RGB images, which are compared to the floorplan for localisation. While our approach is able to use range measurements if available, we demonstrate that they are unnecessary as we can achieve results comparable to state-of-the-art without them.

Abstract We present a novel approach to automatic Sign Language Production using recent developments in Neural Machine Translation (NMT), Generative Adversarial Networks, and motion generation. Our system is capable of producing sign videos from spoken language sentences. Contrary to current approaches that are dependent on heavily annotated data, our approach requires minimal gloss and skeletal level annotations for training. We achieve this by breaking down the task into dedicated sub-processes. We first translate spoken language sentences into sign pose sequences by combining an NMT network with a Motion Graph. The resulting pose information is then used to condition a generative model that produces photo realistic sign language video sequences. This is the first approach to continuous sign video generation that does not use a classical graphical avatar. We evaluate the translation abilities of our approach on the PHOENIX14 T Sign Language Translation dataset. We set a baseline for text-to-gloss translation, reporting a BLEU-4 score of 16.34/15.26 on dev/test sets. We further demonstrate the video generation capabilities of our approach for both multi-signer and high-definition settings qualitatively and quantitatively using broadcast quality assessment metrics.

How do computers and intelligent agents view the world around them? Feature extraction and representation constitutes one the basic building blocks towards answering this question. Traditionally, this has been done with carefully engineered hand-crafted techniques such as HOG, SIFT or ORB. However, there is no "one size fits all" approach that satisfies all requirements. In recent years, the rising popularity of deep learning has resulted in a myriad of end-to-end solutions to many computer vision problems. These approaches, while successful, tend to lack scalability and can't easily exploit information learned by other systems. Instead, we propose SAND features, a dedicated deep learning solution to feature extraction capable of providing hierarchical context information. This is achieved by employing sparse relative labels indicating relationships or similarity of dissimilarity between image locations. The nature of these labels results in an almost infinite set of dissimilar examples to choose from. We demonstrate how the selection of negative examples during training can be used to modify the feature space and vary it's learned properties. To demonstrate the generality of this approach, we apply the proposed features to a multitude of tasks, each requiring different properties. This includes disparity estimation, semantic segmentation, self-localisation and SLAM. In all cases, we show how incorporating SAND features results in better or comparable results to the baseline, whilst requiring little to no additional training.

Accurate extrinsic sensor calibration is essential for both autonomous vehicles and robots. Traditionally this is an involved process requiring calibration targets, known fiducial markers and is generally performed in a lab. Moreover, even a small change in the sensor layout requires recalibration. With the anticipated arrival of consumer autonomous vehicles, there is demand for a system which can do this automatically, after deployment and without specialist human expertise. To solve these limitations, we propose a flexible framework which can estimate extrinsic parameters without an explicit calibration stage, even for sensors with unknown scale. Our first contribution builds upon standard hand-eye calibration by jointly recovering scale. Our second contribution is that our system is made robust to imperfect and degenerate sensor data, by collecting independent sets of poses and automatically selecting those which are most ideal. We show that our approach's robustness is essential for the target scenario. Unlike previous approaches, ours runs in real time and constantly estimates the extrinsic transform. For both an ideal experimental setup and a real use case, comparison against these approaches shows that we outperform the state-of-the-art. Furthermore, we demonstrate that the recovered scale may be applied to the full trajectory, circumventing the need for scale estimation via sensor fusion.

This paper proposes a Hybrid Approximate Representation (HAR) based on unifying several efficient approximations of the generalized reprojection error (which is known as the gold standard for multiview geometry). The HAR is an over-parameterization scheme where the approximation is applied simultaneously in multiple parameter spaces. A joint minimization scheme "HAR-Descent" can then solve the PnP problem efficiently, while remaining robust to approximation errors and local minima. The technique is evaluated extensively, including numerous synthetic benchmark protocols and the real-world data evaluations used in previous works. The proposed technique was found to have runtime complexity comparable to the fastest O(n) techniques, and up to 10 times faster than current state of the art minimization approaches. In addition, the accuracy exceeds that of all 9 previous techniques tested, providing definitive state of the art performance on the benchmarks, across all 90 of the experiments in the paper and supplementary material, which can be found on the Computer Society Digital Library at http://doi.ieeecomputersociety.org/10.1109/TPAMI.2018.2806446.

Sign Language Recognition (SLR) has been an active research field for the last two decades. However, most research to date has considered SLR as a naive gesture recognition problem. SLR seeks to recognize a sequence of continuous signs but neglects the underlying rich grammatical and linguistic structures of sign language that differ from spoken language. In contrast, we introduce the Sign Language Translation (SLT) problem. Here, the objective is to generate spoken language translations from sign language videos, taking into account the different word orders and grammar. We formalize SLT in the framework of Neural Machine Translation (NMT) for both end-to-end and pretrained settings (using expert knowledge). This allows us to jointly learn the spatial representations, the underlying language model, and the mapping between sign and spoken language. To evaluate the performance of Neural SLT, we collected the first publicly available Continuous SLT dataset, RWTH-PHOENIX-Weather 2014T1. It provides spoken language translations and gloss level annotations for German Sign Language videos of weather broadcasts. Our dataset contains over .95M frames with >67K signs from a sign vocabulary of >1K and >99K words from a German vocabulary of >2.8K. We report quantitative and qualitative results for various SLT setups to underpin future research in this newly established field. The upper bound for translation performance is calculated at 19.26 BLEU-4, while our end-to-end frame-level and gloss-level tokenization networks were able to achieve 9.58 and 18.13 respectively.

How does a person work out their location using a floorplan? It is probably safe to say that we do not explicitly measure depths to every visible surface and try to match them against different pose estimates in the floorplan. And yet, this is exactly how most robotic scan-matching algorithms operate. Similarly, we do not extrude the 2D geometry present in the floorplan into 3D and try to align it to the real-world. And yet, this is how most vision-based approaches localise. Humans do the exact opposite. Instead of depth, we use high level semantic cues. Instead of extruding the floorplan up into the third dimension, we collapse the 3D world into a 2D representation. Evidence of this is that many of the floorplans we use in everyday life are not accurate, opting instead for high levels of discriminative landmarks. In this work, we use this insight to present a global localisation approach that relies solely on the semantic labels present in the floorplan and extracted from RGB images. While our approach is able to use range measurements if available, we demonstrate that they are unnecessary as we can achieve results comparable to state-of-the-art without them.

We propose a novel deep learning approach to solve simultaneous alignment and recognition problems (referred to as “Sequence-to-sequence” learning). We decompose the problem into a series of specialised expert systems referred to as SubUNets. The spatio-temporal relationships between these SubUNets are then modelled to solve the task, while remaining trainable end-to-end. The approach mimics human learning and educational techniques, and has a number of significant advantages. SubUNets allow us to inject domain-specific expert knowledge into the system regarding suitable intermediate representations. They also allow us to implicitly perform transfer learning between different interrelated tasks, which also allows us to exploit a wider range of more varied data sources. In our experiments we demonstrate that each of these properties serves to significantly improve the performance of the overarching recognition system, by better constraining the learning problem. The proposed techniques are demonstrated in the challenging domain of sign language recognition. We demonstrate state-of-the-art performance on hand-shape recognition (outperforming previous techniques by more than 30%). Furthermore, we are able to obtain comparable sign recognition rates to previous research, without the need for an alignment step to segment out the signs for recognition.

Reconstruction of 3D environments is a problem that has been widely addressed in the literature. While many approaches exist to perform reconstruction, few of them take an active role in deciding where the next observations should come from. Furthermore, the problem of travelling from the camera's current position to the next, known as pathplanning, usually focuses on minimising path length. This approach is ill-suited for reconstruction applications, where learning about the environment is more valuable than speed of traversal. We present a novel Scenic Route Planner that selects paths which maximise information gain, both in terms of total map coverage and reconstruction accuracy. We also introduce a new type of collaborative behaviour into the planning stage called opportunistic collaboration, which allows sensors to switch between acting as independent Structure from Motion (SfM) agents or as a variable baseline stereo pair. We show that Scenic Planning enables similar performance to state-of-the-art batch approaches using less than 0.00027% of the possible stereo pairs (3% of the views). Comparison against length-based pathplanning approaches show that our approach produces more complete and more accurate maps with fewer frames. Finally, we demonstrate the Scenic Pathplanner's ability to generalise to live scenarios by mounting cameras on autonomous ground-based sensor platforms and exploring an environment.

Action recognition "in the wild" is extremely challenging, particularly when complex 3D actions are projected down to the image plane, losing a great deal of information. The recent growth of 3D data in broadcast content and commercial depth sensors, makes it possible to overcome this. However, there is little work examining the best way to exploit this new modality. In this paper we introduce the Hollywood 3D benchmark, which is the first dataset containing "in the wild" action footage including 3D data. This dataset consists of 650 stereo video clips across 14 action classes, taken from Hollywood movies. We provide stereo calibrations and depth reconstructions for each clip. We also provide an action recognition pipeline, and propose a number of specialised depth-aware techniques including five interest point detectors and three feature descriptors. Extensive tests allow evaluation of different appearance and depth encoding schemes. Our novel techniques exploiting this depth allow us to reach performance levels more than triple those of the best baseline algorithm using only appearance information. The benchmark data, code and calibrations are all made available to the community.

Significant effort has been devoted within the visual tracking community to rapid learning of object properties on the fly. However, state-of-the-art approaches still often fail in cases such as rapid out-of-plane rotation, when the appearance changes suddenly. One of the major contributions of this paper is a radical rethinking of the traditional wisdom of modeling 3D motion as appearance changes during tracking. Instead, 3D motion is modeled as 3D motion. This intuitive but previously unexplored approach provides new possibilities in visual tracking research. First, 3D tracking is more general, as large out-of-plane motion is often fatal for 2D trackers, but helps 3D trackers to build better models. Second, the tracker's internal model of the object can be used in many different applications and it could even become the main motivation, with tracking supporting reconstruction rather than vice versa. This effectively bridges the gap between visual tracking and structure from motion. A new benchmark data set of sequences with extreme out-of-plane rotation is presented and an online leader-board offered to stimulate new research in the relatively underdeveloped area of 3D tracking. The proposed method, provided as a baseline, is capable of successfully tracking these sequences, all of which pose a considerable challenge to 2D trackers (error reduced by 46%).

In this paper, we propose using 3D Convolutional Neural Networks for large scale user-independent continuous gesture recognition. We have trained an end-to-end deep network for continuous gesture recognition (jointly learning both the feature representation and the classifier). The network performs three-dimensional (i.e. space-time) convolutions to extract features related to both the appearance and motion from volumes of color frames. Space-time invariance of the extracted features is encoded via pooling layers. The earlier stages of the network are partially initialized using the work of Tran et al. before being adapted to the task of gesture recognition. An earlier version of the proposed method, which was trained for 11,250 iterations, was submitted to ChaLearn 2016 Continuous Gesture Recognition Challenge and ranked 2nd with the Mean Jaccard Index Score of 0.269235. When the proposed method was further trained for 28,750 iterations, it achieved state-of-the-art performance on the same dataset, yielding a 0.314779 Mean Jaccard Index Score.

Long-term tracking of an object, given only a single instance in an initial frame, remains an open problem. We propose a visual tracking algorithm, robust to many of the difficulties that often occur in real-world scenes. Correspondences of edge-based features are used, to overcome the reliance on the texture of the tracked object and improve invariance to lighting. Furthermore, we address long-term stability, enabling the tracker to recover from drift and to provide redetection following object disappearance or occlusion. The two-module principle is similar to the successful state-of-the-art long-term TLD tracker; however, our approach offers better performance in benchmarks and extends to cases of low-textured objects. This becomes obvious in cases of plain objects with no texture at all, where the edge-based approach proves the most beneficial. We perform several different experiments to validate the proposed method. First, results on short-term sequences show the performance of tracking challenging (low textured and/or transparent) objects that represent failure cases for competing the state-of-the-art approaches. Second, long sequences are tracked, including one of almost 30 000 frames, which, to the best of our knowledge, is the longest tracking sequence reported to date. This tests the redetection and drift resistance properties of the tracker. Finally, we report the results of the proposed tracker on the VOT Challenge 2013 and 2014 data sets as well as on the VTB1.0 benchmark, and we show relative performance of the tracker compared with its competitors. All the results are comparable with the state of the art on sequences with textured objects and superior on non-textured objects. The new annotated sequences are made publicly available.

We present a novel approach to 3D reconstruction which is inspired by the human visual system. This system unifies standard appearance matching and triangulation techniques with higher level reasoning and scene understanding, in order to resolve ambiguities between different interpretations of the scene. The types of reasoning integrated in the approach includes recognising common configurations of surface normals and semantic edges (e.g. convex, concave and occlusion boundaries). We also recognise the coplanar, collinear and symmetric structures which are especially common in man made environments.

Causal relationships can often be found in visual object tracking between the motions of the camera and that of the tracked object. This object motion may be an effect of the camera motion, e.g. an unsteady handheld camera. But it may also be the cause, e.g. the cameraman framing the object. In this paper we explore these relationships, and provide statistical tools to detect and quantify them, these are based on transfer entropy and stem from information theory. The relationships are then exploited to make predictions about the object location. The approach is shown to be an excellent measure for describing such relationships. On the VOT2013 dataset the prediction accuracy is increased by 62 % over the best non-causal predictor. We show that the location predictions are robust to camera shake and sudden motion, which is invaluable for any tracking algorithm and demonstrate this by applying causal prediction to two state-of-the-art trackers. Both of them benefit, Struck gaining a 7 % accuracy and 22 % robustness increase on the VTB1.1 benchmark, becoming the new state-of-the-art.

In this paper, an algorithm is presented for estimating scene flow, which is a richer, 3D analog of optical flow. The approach operates orders of magnitude faster than alternative techniques and is well suited to further performance gains through parallelized implementation. The algorithm employs multiple hypotheses to deal with motion ambiguities, rather than the traditional smoothness constraints, removing oversmoothing errors and providing significant performance improvements on benchmark data, over the previous state of the art. The approach is flexible and capable of operating with any combination of appearance and/or depth sensors, in any setup, simultaneously estimating the structure and motion if necessary. Additionally, the algorithm propagates information over time to resolve ambiguities, rather than performing an isolated estimation at each frame, as in contemporary approaches. Approaches to smoothing the motion field without sacrificing the benefits of multiple hypotheses are explored, and a probabilistic approach to occlusion estimation is demonstrated, leading to 10 and 15 percent improved performance, respectively. Finally, a data-driven tracking approach is described, and used to estimate the 3D trajectories of hands during sign language, without the need to model complex appearance variations at each viewpoint.

Action recognition in unconstrained situations is a difficult task, suffering from massive intra-class variations. It is made even more challenging when complex 3D actions are projected down to the image plane, losing a great deal of information. The recent emergence of 3D data, both in broadcast content, and commercial depth sensors, provides the possibility to overcome this issue. This paper presents a new dataset, for benchmarking action recognition algorithms in natural environments, while making use of 3D information. The dataset contains around 650 video clips, across 14 classes. In addition, two state of the art action recognition algorithms are extended to make use of the 3D data, and five new interest point detection strategies are also proposed, that extend to the 3D data. Our evaluation compares all 4 feature descriptors, using 7 different types of interest point, over a variety of threshold levels, for the Hollywood3D dataset. We make the dataset including stereo video, estimated depth maps and all code required to reproduce the benchmark results, available to the wider community.

The motion field of a scene can be used for object segmentation and to provide features for classification tasks like action recognition. Scene flow is the full 3D motion field of the scene, and is more difficult to estimate than it's 2D counterpart, optical flow. Current approaches use a smoothness cost for regularisation, which tends to over-smooth at object boundaries. This paper presents a novel formulation for scene flow estimation, a collection of moving points in 3D space, modelled using a particle filter that supports multiple hypotheses and does not oversmooth the motion field. In addition, this paper is the first to address scene flow estimation, while making use of modern depth sensors and monocular appearance images, rather than traditional multi-viewpoint rigs. The algorithm is applied to an existing scene flow dataset, where it achieves comparable results to approaches utilising multiple views, while taking a fraction of the time.