Image goal navigation requires two different skills: firstly, core navigation skills, including the detection of free space and obstacles, and taking decisions based on an internal representation; and secondly, computing directional information by comparing visual observations to the goal image. Current state-of-the-art methods either rely on dedicated image-matching, or pre-training of computer vision modules on relative pose estimation. In this paper, we study whether this task can be efficiently solved with end-to-end training of full agents with RL, as has been claimed by recent work. A positive answer would have impact beyond Embodied AI and allow training of relative pose estimation from reward for navigation alone. In this large experimental study we investigate the effect of architectural choices like late fusion, channel stacking, space-to-depth projections and cross-attention, and their role in the emergence of relative pose estimators from navigation training. We show that the success of recent methods is influenced up to a certain extent by simulator settings, leading to shortcuts in simulation. However, we also show that these capabilities can be transferred to more realistic setting, up to some extent. We also find evidence for correlations between navigation performance and probed (emerging) relative pose estimation performance, an important sub skill.

Progress in Embodied AI has made it possible for end-to-end-trained agents to navigate in photo-realistic environments with high-level reasoning and zero-shot or language-conditioned behavior, but benchmarks are still dominated by simulation. In this work, we focus on the fine-grained behavior of fast-moving real robots and present a large-scale experimental study involving 262 navigation episodes in a real environment with a physical robot, where we analyze the type of reasoning emerging from end-to-end training. In particular, we study the presence of realistic dynamics which the agent learned for open-loop forecasting, and their interplay with sensing. We analyze the way the agent uses latent memory to hold elements of the scene structure and information gathered during exploration. We probe the planning capabilities of the agent, and find in its memory evidence for somewhat precise plans over a limited horizon. Furthermore, we show in a post-hoc analysis that the value function learned by the agent relates to long-term planning. Put together, our experiments paint a new picture on how using tools from computer vision and sequential decision making have led to new capabilities in robotics and control. An interactive tool is available [here].

One key aspect of spatially aware robots is the ability to "find their bearings", ie. to correctly situate themselves or previously seen spaces. In this work, we focus on this particular scenario of continuous robotics operations, where information observed before an actual episode start is exploited to optimize efficiency. We introduce a new model, "Kinaema" and agent, capable of integrating a stream of visual observations while moving in a potentially large scene, and upon request, processing a query image and predicting the relative position of the shown space with respect to its current position. Our model does not explicitly store an observation history, therefore does not have hard constraints on context length. It maintains an implicit latent memory, which is updated by a transformer in a recurrent way, compressing the history of sensor readings into a compact representation. We evaluate the impact of this model in a new downstream task we call "Mem-Nav", targeting continuous robotics operations. We show that our large-capacity recurrent model maintains a useful representation of the scene, navigates to goals observed before the actual episode start, and is computationally efficient, in particular compared to classical transformers with attention over an observation history.

Bird’s-eye view (BEV) maps are an important geometrically structured representation widely used in robotics, in particular self-driving vehicles and terrestrial robots. Existing algorithms either require depth information for the geometric projection, which is not always reliably available, or are trained end-to-end in a fully supervised way to map visual first-person observations to BEV representation, and are therefore restricted to the output modality they have been trained for. In contrast, we propose a new model capable of performing zero-shot projections of any modality available in a first person view to the corresponding BEV map. This is achieved by disentangling the geometric inverse perspective projection from the modality transformation, e.g. RGB to occupancy. The method is general and we showcase experiments projecting to BEV three different modalities: semantic segmentation, motion vectors and object bounding boxes detected in first person. We experimentally show that the model outperforms competing methods, in particular the widely used baseline resorting to monocular depth estimation.

In the context of autonomous navigation of terrestrial robots, the creation of realistic models for agent dynamics and sensing is a widespread habit in the robotics literature and in commercial applications, where they are used for model based control and/or for localization and mapping. The more recent Embodied AI literature, on the other hand, focuses on modular or end-to-end agents trained in simulators like Habitat or AI-Thor, where the emphasis is put on photo-realistic rendering and scene diversity, but high-fidelity robot motion is assigned a less privileged role. The resulting sim2real gap significantly impacts transfer of the trained models to real robotic platforms. In this work we explore end-to-end training of agents in simulation in settings which minimize the sim2real gap both, in sensing and in actuation. Our agent directly predicts (discretized) velocity commands, which are maintained through closed-loop control in the real robot. The behavior of the real robot (including the underlying low-level controller) is identified and simulated in a modified Habitat simulator. Noise models for odometry and localization further contribute in lowering the sim2real gap. We evaluate on real navigation scenarios, explore different localization and point goal calculation methods and report significant gains in performance and robustness compared to prior work.

Successfully addressing a wide variety of tasks is a core ability of autonomous agents, requiring flexibly adapting the underlying decision-making strategies and, as we ar-gue in this work, also adapting the perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, we adapt pretrained large vision models conditioned on specific downstream tasks in the context of multitask policy learning. We introduce task-conditioned adapters that do not require finetuning any pretrained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations.

Most recent work in goal oriented visual navigation resorts to large-scale machine learning in simulated environments. The main challenge lies in learning compact representations generalizable to unseen environments and in learning high-capacity perception modules capable of reasoning on high-dimensional input. The latter is particularly difficult when the goal is not given as a category ("ObjectNav") but as an exemplar image ("ImageNav"), as the perception module needs to learn a comparison strategy requiring to solve an underlying visual correspondence problem. This has been shown to be difficult from reward alone or with standard auxiliary tasks. We address this problem through a sequence of two pretext tasks, which serve as a prior for what we argue is one of the main bottleneck in perception, extremely wide-baseline relative pose estimation and visibility prediction in complex scenes. The first pretext task, cross-view completion is a proxy for the underlying visual correspondence problem, while the second task addresses goal detection and finding directly. We propose a new dual encoder with a large-capacity binocular ViT model and show that correspondence solutions naturally emerge from the training signals. Experiments show significant improvements and SOTA performance on the two benchmarks, ImageNav and the Instance-ImageNav variant, where camera intrinsics and height differ between observation and goal.

Agents navigating in 3D environments require some form of memory, which should hold a compact and actionable representation of the history of observations useful for decision taking and planning. In most end-to-end learning approaches the representation is latent and usually does not have a clearly defined interpretation, whereas classical robotics addresses this with scene reconstruction resulting in some form of map, usually estimated with geometry and sensor models and/or learning. In this work we propose to learn an actionable representation of the scene independently of the targeted downstream task and without explicitly optimizing reconstruction. The learned representation is optimized by a blind auxiliary agent trained to navigate with it on multiple short sub episodes branching out from a waypoint and, most importantly, without any direct visual observation. We argue and show that the blindness property is important and forces the (trained) latent representation to be the only means for planning. With probing experiments we show that the learned representation optimizes navigability and not reconstruction. On downstream tasks we show that it is robust to changes in distribution, in particular the sim2real gap, which we evaluate with a real physical robot in a real office building, significantly improving performance.

Modern techniques for physical simulations rely on numerical schemes and mesh-refinement methods to address trade-offs between precision and complexity, but these handcrafted solutions are tedious and require high computational power. Data-driven methods based on large-scale machine learning promise high adaptivity by integrating long-range dependencies more directly and efficiently. In this work, we focus on computational fluid dynamics and address the shortcomings of a large part of the literature, which are based on fixed support for computations and predictions in the form of regular or irregular grids. We propose a novel setup to perform predictions in a continuous spatial and temporal domain while being trained on sparse observations. We formulate the task as a double observation problem and propose a solution with two interlinked dynamical systems defined on, respectively, the sparse positions and the continuous domain, which allows to forecast and interpolate a solution from the initial condition. Our practical implementation involves recurrent GNNs and a spatio-temporal attention observer capable of interpolating the solution at arbitrary locations. Our model not only generalizes to new initial conditions (as standard auto-regressive models do) but also performs evaluation at arbitrary space and time locations. We evaluate on three standard datasets in fluid dynamics and compare to strong baselines, which are outperformed in classical settings and the extended new task requiring continuous predictions.

Implicit representations such as Neural Radiance Fields (NeRF) allow to map color, density and semantics in a 3D scene through a continuous neural function. However, these models typically require manual and careful human data collection for training. This paper addresses the problem of active exploration for autonomous NeRF construction. We study how an agent can learn to efficiently explore an unknown 3D environment so that the data collected during autonomous exploration enables the learning of a high-quality neural implicit map representation. The quality of the learned representation is evaluated on four robotics-related downstream tasks: classical viewpoint rendering, map reconstruction, planning, and pose refinement. We compare the impact of different exploration strategies including frontier-based and learning-based approaches (end-to-end and modular) with different reward functions tailored to this problem. Empirical results show that NeRFs can be trained on actively collected data using just a single episode of experience in an unseen environment and that AutoNeRF, a modular exploration policy trained with reinforcement learning, enables obtaining a higher-quality NeRF for the considered downstream robotic tasks. Finally, we show that with AutoNeRF an agent can be deployed to a previously unknown scene and then automatically improve its navigation performance by adapting to the scene through a cycle of exploration, reconstruction, and policy finetuning.

Understanding and mapping a new environment are core abilities of any autonomously navigating agent. While classical robotics usually estimates maps in a stand-alone manner with SLAM variants, which maintain a topological or metric representation, end-to-end learning of navigation keeps some form of memory in a neural network. Networks are typically imbued with inductive biases, which can range from vectorial representations to birds-eye metric tensors or topological structures. In this work, we propose to structure neural networks with two neural implicit representations, which are learned dynamically during each episode and map the content of the scene: (i) the Semantic Finder predicts the position of a previously seen queried object; (ii) the Occupancy and Exploration Implicit Representation encapsulates information about explored area and obstacles, and is queried with a novel global read mechanism which directly maps from function space to a usable embedding space. Both representations are leveraged by an agent trained with Reinforcement Learning (RL) and learned online during each episode. We evaluate the agent on Multi-Object Navigation and show the high impact of using neural implicit representations as a memory source.

Navigation has been classically solved in robotics through the combination of SLAM and planning. More recently, beyond waypoint planning, problems involving significant components of (visual) high-level reasoning have been explored in simulated environments, mostly addressed with large-scale machine learning, in particular RL, offline-RL or imitation learning. These methods require the agent to learn various skills like local planning, mapping objects and querying the learned spatial representations. In contrast to simpler tasks like waypoint planning (PointGoal), for these more complex tasks the current state-of-the-art models have been thoroughly evaluated in simulation but, to our best knowledge, not yet in real environments. In this work we focus on sim2real transfer. We target the challenging Multi-Object Navigation (Multi-ON) task [41] and port it to a physical environment containing real replicas of the originally virtual Multi-ON objects. We introduce a hybrid navigation method, which decomposes the problem into two different skills: (1) waypoint navigation is addressed with classical SLAM combined with a symbolic planner, whereas (2) exploration, semantic mapping and goal retrieval are dealt with deep neural networks trained with a combination of supervised learning and RL. We show the advantages of this approach compared to end-to-end methods both in simulation and a real environment and outperform the SOTA for this task [28].

Navigation of terrestrial robots is typically addressed either with localization and mapping (SLAM) followed by classical planning on the dynamically created maps, or by machine learning (ML), often through end-to-end training with reinforcement learning (RL) or imitation learning (IL). Recently, modular designs have achieved promising results, and hybrid algorithms that combine ML with classical planning have been proposed. Existing methods implement these combinations with handcrafted functions, which cannot fully exploit the complementary nature of the policies and the complex regularities between scene structure and planning performance. Our work builds on the hypothesis that the strengths and weaknesses of neural planners and classical planners follow some regularities, which can be learned from training data, in particular from interactions. This is grounded on the assumption that, both, trained planners and the mapping algorithms underlying classical planning are subject to failure cases depending on the semantics of the scene and that this dependence is learnable: for instance, certain areas, objects or scene structures can be reconstructed easier than others. We propose a hierarchical method composed of a high-level planner dynamically switching between a classical and a neural planner. We fully train all neural policies in simulation and evaluate the method in both simulation and real experiments with a LoCoBot robot, showing significant gains in performance, in particular in the real environment. We also qualitatively conjecture on the nature of data regularities exploited by the high-level planner.

Visual navigation by mobile robots is classically tackled through SLAM plus optimal planning, and more recently through end-to-end training of policies implemented as deep networks. While the former are often limited to waypoint planning, but have proven their efficiency even on real physical environments, the latter solutions are most frequently employed in simulation, but have been shown to be able learn more complex visual reasoning, involving complex semantic regularities. Navigation by real robots in physical environments is still an open problem. End-to-end training approaches have been thoroughly tested in simulation only, with experiments involving real robots being restricted to rare performance evaluations in simplified laboratory conditions. In this work we present an in-depth study of the performance and reasoning capacities of real physical agents, trained in simulation and deployed to two different physical environments. Beyond benchmarking, we provide insights into the generalization capabilities of different agents training in different conditions. We visualize sensor usage and the importance of the different types of signals. We show, that for the PointGoal task, an agent pre-trained on wide variety of tasks and fine-tuned on a simulated version of the target environment can reach competitive performance without modelling any sim2real transfer, i.e. by deploying the trained agent directly from simulation to a real physical robot.

In the context of visual navigation, the capacity to map a novel environment is necessary for an agent to exploit its observation history in the considered place and efficiently reach known goals. This ability can be associated with spatial rea-soning, where an agent is able to perceive spatial relationships and regularities, and discover object characteristics. Recent work introduces learnable policies parametrized by deep neural networks and trained with Reinforcement Learning (RL). In classical RL setups, the capacity to map and reason spatially is learned end-to-end, from reward alone. In this setting, we introduce supplementary supervision in the form of auxiliary tasks designed to favor the emergence of spatial perception capabilities in agents trained for a goal-reaching downstream objective. We show that learning to estimate metrics quantifying the spatial relationships between an agent at a given location and a goal to reach has a high positive impact in Multi-Object Navigation settings. Our method significantly improves the performance of different baseline agents, that either build an explicit or implicit representation of the environment, even matching the performance of incomparable oracle agents taking ground-truth maps as input. A learning-based agent from the literature trained with the proposed auxiliary losses was the winning entry to the Multi-Object Navigation Challenge, part of the CVPR 2021 Embodied AI Workshop.

Visual Question Answering systems target answering open-ended textual questions given input images. They are a testbed for learning high-level reasoning with a primary use in HCI, for instance assistance for the visually impaired. Recent research has shown that state-of-the-art models tend to produce answers exploiting biases and shortcuts in the training data, and sometimes do not even look at the input image, instead of performing the required reasoning steps. We present VisQA, a visual analytics tool that explores this question of reasoning vs. bias exploitation. It exposes the key element of state-of-the-art neural models - attention maps in transformers. Our working hypothesis is that reasoning steps leading to model predictions are observable from attention distributions, which are particularly useful for visualization. The design process of VisQA was motivated by well-known bias examples from the fields of deep learning and vision-language reasoning and evaluated in two ways. First, as a result of a collaboration of three fields, machine learning, vision and language reasoning, and data analytics, the work lead to a better understanding of bias exploitation of neural models for VQA, which eventually resulted in an impact on its design and training through the proposition of a method for the transfer of reasoning patterns from an oracle model. Second, we also report on the design of VisQA, and a goal-oriented evaluation of VisQA targeting the analysis of a model decision process from multiple experts, providing evidence that it makes the inner workings of models accessible to users.

In this paper we introduce a Transformer-based approach to video object segmentation (VOS). To address compounding error and scalability issues of prior work, we propose a scalable, end-to-end method for VOS called Sparse Spatiotemporal Transformers (SST). SST extracts per-pixel representations for each object in a video using sparse attention over spatiotemporal features. Our attention-based formulation for VOS allows a model to learn to attend over a history of multiple frames and provides suitable inductive bias for performing correspondence-like computations necessary for solving motion segmentation. We demonstrate the effectiveness of attention-based over recurrent networks in the spatiotemporal domain. Our method achieves competitive results on YouTube-VOS and DAVIS 2017 with improved scalability and robustness to occlusions compared with the state of the art. Code is available at https://github.com/dukebw/SSTVOS.

Models for Visual Question Answering (VQA) are notorious for their tendency to rely on dataset biases, as the large and unbalanced diversity of questions and concepts involved and tends to prevent models from learning to "reason", leading them to perform "educated guesses" instead. In this paper, we claim that the standard evaluation metric, which consists in measuring the overall in-domain accuracy, is misleading. Since questions and concepts are unbalanced, this tends to favor models which exploit subtle training set statistics. Alternatively, naively introducing artificial distribution shifts between train and test splits is also not completely satisfying. First, the shifts do not reflect real-world tendencies, resulting in unsuitable models; second, since the shifts are handcrafted, trained models are specifically designed for this particular setting, and do not generalize to other configurations. We propose the GQAOOD benchmark designed to overcome these concerns: we measure and compare accuracy over both rare and frequent question-answer pairs, and argue that the former is better suited to the evaluation of reasoning abilities, which we experimentally validate with models trained to more or less exploit biases. In a large-scale study involving 7 VQA models and 3 bias reduction techniques, we also experimentally demonstrate that these models fail to address questions involving infrequent concepts and provide recommendations for future directions of research.

Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available1. We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.

Methods for Visual Question Anwering (VQA) are notorious for leveraging dataset biases rather than performing reasoning, hindering generalization. It has been recently shown that better reasoning patterns emerge in attention layers of a state-of-the-art VQA model when they are trained on perfect (oracle) visual inputs. This provides evidence that deep neural networks can learn to reason when training conditions are favorable enough. However, transferring this learned knowledge to deployable models is a challenge, as much of it is lost during the transfer.We propose a method for knowledge transfer based on a regularization term in our loss function, supervising the sequence of required reasoning operations.We provide a theoretical analysis based on PAC-learning, showing that such program prediction can lead to decreased sample complexity under mild hypotheses. We also demonstrate the effectiveness of this approach experimentally on the GQA dataset and show its complementarity to BERT-like self-supervised pre-training.

The large adoption of the self-attention (i.e. transformer model) and BERT-like training principles has recently resulted in a number of high performing models on a large panoply of vision-and-language problems (such as Visual Question Answering (VQA), image retrieval, etc.). In this paper we claim that these State-Of-The-Art (SOTA) approaches perform reasonably well in structuring information inside a single modality but, despite their impressive performances , they tend to struggle to identify fine-grained inter-modality relationships. Indeed, such relations are frequently assumed to be implicitly learned during training from application-specific losses, mostly cross-entropy for classification. While most recent works provide inductive bias for inter-modality relationships via cross attention modules, in this work, we demonstrate (1) that the latter assumption does not hold, i.e. modality alignment does not necessarily emerge automatically, and (2) that adding weak supervision for alignment between visual objects and words improves the quality of the learned models on tasks requiring reasoning. In particular , we integrate an object-word alignment loss into SOTA vision-language reasoning models and evaluate it on two tasks VQA and Language-driven Comparison of Images. We show that the proposed fine-grained inter-modality supervision significantly improves performance on both tasks. In particular, this new learning signal allows obtaining SOTA-level performances on GQA dataset (VQA task) with pre-trained models without finetuning on the task, and a new SOTA on NLVR2 dataset (Language-driven Comparison of Images). Finally, we also illustrate the impact of the contribution on the models reasoning by visualizing attention distributions.

An important goal of research in Deep Reinforcement Learning in mobile robotics is to train agents capable of solving complex tasks, which require a high level of scene understanding and reasoning from an egocentric perspective. When trained from simulations, optimal environments should satisfy a currently unobtainable combination of high-fidelity photographic observations, massive amounts of different environment configurations and fast simulation speeds. In this paper we argue that research on training agents capable of complex reasoning can be simplified by decoupling from the requirement of high fidelity photographic observations. We present a suite of tasks requiring complex reasoning and exploration in continuous, partially observable 3D environments. The objective is to provide challenging scenarios and a robust baseline agent architecture that can be trained on mid-range consumer hardware in under 24h. Our scenarios combine two key advantages: (i) they are based on a simple but highly efficient 3D environment (ViZDoom) which allows high speed simulation (12000fps); (ii) the scenarios provide the user with a range of difficulty settings, in order to identify the limitations of current state of the art algorithms and network architectures. We aim to increase accessibility to the field of Deep-RL by providing baselines for challenging scenarios where new ideas can be iterated on quickly. We argue that the community should be able to address challenging problems in reasoning of mobile agents without the need for a large compute infrastructure. Code for the generation of scenarios and training of baselines is available online at the following repository1.1https://github.com/edbeeching/3d_control_deep_rl.

Abstract We present DRLViz, a visual analytics interface to interpret the internal memory of an agent (e.g. a robot) trained using deep reinforcement learning. This memory is composed of large temporal vectors updated when the agent moves in an environment and is not trivial to understand due to the number of dimensions, dependencies to past vectors, spatial/temporal correlations, and co‐correlation between dimensions. It is often referred to as a black box as only inputs (images) and outputs (actions) are intelligible for humans. Using DRLViz, experts are assisted to interpret decisions using memory reduction interactions, and to investigate the role of parts of the memory when errors have been made (e.g. wrong direction). We report on DRLViz applied in the context of video games simulators (ViZDoom) for a navigation scenario with item gathering tasks. We also report on experts evaluation using DRLViz, and applicability of DRLViz to other scenarios and navigation problems beyond simulation games, as well as its contribution to black box models interpretability and explain‐ability in the field of visual analytics.

We propose a method for human activity recognition from RGB data that does not rely on any pose information during test time, and does not explicitly calculate pose information internally. Instead, a visual attention module learns to predict glimpse sequences in each frame. These glimpses correspond to interest points in the scene that are relevant to the classified activities. No spatial coherence is forced on the glimpse locations, which gives the attention module liberty to explore different points at each frame and better optimize the process of scrutinizing visual information. Tracking and sequentially integrating this kind of unstructured data is a challenge, which we address by separating the set of glimpses from a set of recurrent tracking/recognition workers. These workers receive glimpses, jointly performing subsequent motion tracking and activity prediction. The glimpses are soft-assigned to the workers, optimizing coherence of the assignments in space, time and feature space using an external memory module. No hard decisions are taken, i.e. each glimpse point is assigned to all existing workers, albeit with different importance. Our methods outperform the state-of-the-art on the largest human activity recognition dataset available to-date, NTU RGB+D, and on the Northwestern-UCLA Multiview Action 3D Dataset.

We propose a fully automatic method for learning gestures on big touch devices in a potentially multi-user context. The goal is to learn general models capable of adapting to different gestures, user styles and hardware variations (e.g. device sizes, sampling frequencies and regularities). Based on deep neural networks, our method features a novel dynamic sampling and temporal normalization component, transforming variable length gestures into fixed length representations while preserving finger/surface contact transitions, that is, the topology of the signal. This sequential representation is then processed with a convolutional model capable, unlike recurrent networks, of learning hierarchical representations with different levels of abstraction. To demonstrate the interest of the proposed method, we introduce a new touch gestures dataset with 6591 gestures performed by 27 people, which is, up to our knowledge, the first of its kind: a publicly available multi-touch gesture dataset for interaction. We also tested our method on a standard dataset of symbolic touch gesture recognition, the MMG dataset, outperforming the state of the art and reporting close to perfect performance.

We propose a method to estimate the semantic grid for an autonomous vehicle. The semantic grid is a 2D bird's eye view map where the grid cells contain semantic characteristics such as road, car, pedestrian, signage, etc. We obtain the semantic grid by fusing the semantic segmentation information and an occupancy grid computed by using a Bayesian filter technique. To compute the semantic information from a monocular RGB image, we integrate segmentation deep neural networks into our model. We use a deep neural network to learn the relation between the semantic information and the occupancy grid which can be trained end-to-end extending our previous work on semantic grids. Furthermore, we investigate the effect of using a conditional random field to refine the results. Finally, we test our method on two datasets and compare different architecture types for semantic segmentation. We perform the experiments on KITTI dataset and Inria-Chroma dataset.

In an autonomous vehicle setting, we propose a method for the estimation of a semantic grid, i.e. a bird's eye grid centered on the car's position and aligned with its driving direction, which contains high-level semantic information about the environment and its actors. Each grid cell contains a semantic label with divers classes, as for instance {Road, Vegetation, Building, Pedestrian, Car...}. We propose a hybrid approach, which combines the advantages of two different methodologies: we use Deep Learning to perform semantic segmentation on monocular RGB images with supervised learning from labeled groundtruth data. We combine these segmentations with occupancy grids calculated from LIDAR data using a generative Bayesian particle filter. The fusion itself is carried out with a deep neural network, which learns to integrate geometric information from the LIDAR with semantic information from the RGB data. We tested our method on two datasets, namely the KITTI dataset, which is publicly available and widely used, and our own dataset obtained with our own platform, equipped with a LIDAR and various sensors. We largely outperform baselines which calculate the semantic grid either from the RGB image alone or from LIDAR output alone, showing the interest of this hybrid approach.

We propose a new method for human pose estimation which leverages information from multiple views to impose a strong prior on articulated pose. The novelty of the method concerns the types of coherence modelled. Consistency is maximised over the different views through different terms modelling classical geometric information (coherence of the resulting poses) as well as appearance information which is modelled as latent variables in the global energy function. Moreover, adequacy of each view is assessed and their contributions are adjusted accordingly. Experiments on the HumanEva and Utrecht multi‐person motion datasets show that the proposed method significantly decreases the estimation error compared to single‐view results.

Model selection methods based on stochastic regularization have been widely used in deep learning due to their simplicity and effectiveness. The well-known Dropout method treats all units, visible or hidden, in the same way, thus ignoring any a priori information related to grouping or structure. Such structure is present in multi-modal learning applications such as affect analysis and gesture recognition, where subsets of units may correspond to individual modalities. Here we describe Modout, a model selection method based on stochastic regularization, which is particularly useful in the multi-modal setting. Different from other forms of stochastic regularization, it is capable of learning whether or when to fuse two modalities in a layer, which is usually considered to be an architectural hyper-parameter by deep learning researchers and practitioners. Modout is evaluated on two real multi-modal datasets. The results indicate improved performance compared to other forms of stochastic regularization. The result on the Montalbano dataset shows that learning a fusion structure by Modout is on par with a state-of-the-art carefully designed architecture.

Text line detection and localization is a crucial step for full page document analysis, but still suffers from heterogeneity of real life documents. In this paper, we present a new approach for full page text recognition. Localization of the text lines is based on regressions with Fully Convolutional Neural Networks and Multidimensional Long Short-Term Memory as contextual layers. In order to increase the efficiency of this localization method, only the position of the left side of the text lines are predicted. The text recognizer is then in charge of predicting the end of the text to recognize. This method has shown good results for full page text recognition on the highly heterogeneous Maurdor dataset.

Authoring virtual terrains presents a challenge and there is a strong need for authoring tools able to create realistic terrains with simple user-inputs and with high user control. We propose an example-based authoring pipeline that uses a set of terrain synthesizers dedicated to specific tasks. Each terrain synthesizer is a Conditional Generative Adversarial Network trained by using real-world terrains and their sketched counterparts. The training sets are built automatically with a view that the terrain synthesizers learn the generation from features that are easy to sketch. During the authoring process, the artist first creates a rough sketch of the main terrain features, such as rivers, valleys and ridges, and the algorithm automatically synthesizes a terrain corresponding to the sketch using the learned features of the training samples. Moreover, an erosion synthesizer can also generate terrain evolution by erosion at a very low computational cost. Our framework allows for an easy terrain authoring and provides a high level of realism for a minimum sketch cost. We show various examples of terrain synthesis created by experienced as well as inexperienced users who are able to design a vast variety of complex terrains in a very short time.

Text line detection and localisation is a crucial step for full page document analysis, but still suffers from heterogeneity of real life documents. In this paper, we present a novel approach for text line localisation based on Convolutional Neural Networks and Multidimensional Long Short-Term Memory cells as a regressor in order to predict the coordinates of the text line bounding boxes directly from the pixel values. Targeting typically large images in document image analysis, we propose a new model using weight sharing over local blocks. We compare two strategies: directly predicting the four coordinates or predicting lower-left and upper-right points separately followed by matching. We evaluate our work on the highly unconstrained Maurdor dataset and show that our method outperforms both other machine learning and image processing methods.

Color constancy is the ability of the human visual system to perceive constant colors for a surface despite changes in the spectrum of the illumination. In computer vision, the main approach consists in estimating the illuminant color and then to remove its impact on the color of the objects. Many image processing algorithms have been proposed to tackle this problem automatically. However, most of these approaches are handcrafted and mostly rely on strong empirical assumptions, e.g., that the average reflectance in a scene is gray. State-of-the-art approaches can perform very well on some given datasets but poorly adapt on some others. In this paper, we have investigated how neural networks-based approaches can be used to deal with the color constancy problem. We have proposed a new network architecture based on existing successful hand-crafted approaches and a large number of improvements to tackle this problem by learning a suitable deep model. We show our results on most of the standard benchmarks used in the color constancy domain.

We present a method for gesture detection and localisation based on multi-scale and multi-modal deep learning. Each visual modality captures spatial information at a particular spatial scale (such as motion of the upper body or a hand), and the whole system operates at three temporal scales. Key to our technique is a training strategy which exploits: i) careful initialization of individual modalities; and ii) gradual fusion involving random dropping of separate channels (dubbed ModDrop) for learning cross-modality correlations while preserving uniqueness of each modality-specific representation. We present experiments on the ChaLearn 2014 Looking at People Challenge gesture recognition track, in which we placed first out of 17 teams. Fusing multiple modalities at several spatial and temporal scales leads to a significant increase in recognition rates, allowing the model to compensate for errors of the individual classifiers as well as noise in the separate channels. Furthermore, the proposed ModDrop training technique ensures robustness of the classifier to missing signals in one or several channels to produce meaningful predictions from any number of available modalities. In addition, we demonstrate the applicability of the proposed fusion scheme to modalities of arbitrary nature by experiments on the same dataset augmented with audio.

The detection of text lines, as a first processing step, is critical in all text recognition systems. State-of-the-art methods to locate lines of text are based on handcrafted heuristics fine-tuned by the image processing community's experience. They succeed under certain constraints; for instance the background has to be roughly uniform. We propose to use more “agnostic” Machine Learning-based approaches to address text line location. The main motivation is to be able to process either damaged documents, or flows of documents with a high variety of layouts and other characteristics. A new method is presented in this work, inspired by the latest generation of optical models used for text recognition, namely Recurrent Neural Networks. As these models are sequential, a column of text lines in our application plays here the same role as a line of characters in more traditional text recognition settings. A key advantage of the proposed method over other data-driven approaches is that compiling a training dataset does not require labeling line boundaries: only the number of lines are required for each paragraph. Experimental results show that our approach gives similar or better results than traditional handcrafted approaches, with little engineering efforts and less hyper-parameter tuning.

We propose a new method for activity recognition based on a view independent representation of human motion. Robust 3D volume motion templates (VMTs) are calculated from tracklets. View independence is achieved through a rotation with respect to a canonical orientation. From this volumes, features based on 3D gradients are extracted, projected to a codebook and pooled into a bags-of-words model classified with an SVM classifier. Experiments show that the method outperforms the original HoG3D method.

Modeling multimodal perception-action loops in face-to-face interactions is a crucial step in the process of building sensory-motor behaviors for social robots or users-aware Embodied Conversational Agents (ECA). In this paper, we compare trainable behavioral models based on sequential models (HMMs) and classifiers (SVMs and Decision Trees) inherently inappropriate to model sequential aspects. These models aim at giving pertinent perception/action skills for robots in order to generate optimal actions given the perceived actions of others and joint goals. We applied these models to parallel speech and gaze data collected from interacting dyads. The challenge was to predict the gaze of one subject given the gaze of the interlocutor and the voice activity of both. We show that Incremental Discrete HMM (IDHMM) generally outperforms classifiers and that injecting input context in the modeling process significantly improves the performances of all algorithms.

International audience

In this paper we tackle the problem of detecting individual human actions in video sequences. While the most successful methods are based on local features, which proved that they can deal with changes in background, scale and illumination, most existing methods have two main shortcomings: first, they are mainly based on the individual power of spatio-temporal interest points (STIP), and therefore ignore the spatio-temporal relationships between them. Second, these methods mainly focus on direct classification techniques to classify the human activities, as opposed to detection and localization. In order to overcome these limitations, we propose a new approach, which is based on a graph matching algorithm for activity recognition. In contrast to most previous methods which classify entire video sequences, we design a video matching method from two sets of ST-points for human activity recognition. First, points are extracted, and a hyper graphs are constructed from them, i.e. graphs with edges involving more than 2 nodes (3 in our case). The activity recognition problem is then transformed into a problem of finding instances of model graphs in the scene graph. By matching local features instead of classifying entire sequences, our method is able to detect multiple different activities which occur simultaneously in a video sequence. Experiments on two standard datasets demonstrate that our method is comparable to the existing techniques on classification, and that it can, additionally, detect and localize activities.

Existing action recognition approaches mainly rely on the discriminative power of individual local descriptors extracted from spatio-temporal interest points (STIP), while the geometric relationships among the local features are ignored. This paper presents new features, called pairwise features (PWF), which encode both the appearance and the spatio-temporal relations of the local features for action recognition. First STIPs are extracted, then PWFs are constructed by grouping pairs of STIPs which are both close in space and close in time. We propose a combination of two codebooks for video representation. Experiments on two standard human action datasets: the KTH dataset and the Weizmann dataset show that the proposed approach outperforms most existing methods.

GMM based algorithms have become the de facto standard for background subtraction in video sequences, mainly because of their ability to track multiple background distributions, which allows them to handle complex scenes including moving trees, flags moving in the wind etc. However, it is not always easy to determine which distributions of the mixture belong to the background and which distributions belong to the foreground, which disturbs the results of the labeling process for each pixel. In this work we tackle this problem by taking the labeling decision together for all pixels of several consecutive frames minimizing a global energy function taking into account spatial and temporal relationships. A discrete approximative optical-flow like motion model is integrated into the energy function and solved with Ishikawa's convex graph cuts algorithm.

We present a new method for blind document bleed-through removal based on separate Markov Random Field (MRF) regularization for the recto and for the verso side, where separate priors are derived from the full graph. The segmentation algorithm is based on Bayesian Maximum a Posteriori (MAP) estimation. The advantages of this separate approach are the adaptation of the prior to the contents creation process (e.g., superimposing two handwritten pages), and the improvement of the estimation of the recto pixels through an estimation of the verso pixels covered by recto pixels; moreover, the formulation as a binary labeling problem with two hidden labels per pixels naturally leads to an efficient optimization method based on the minimum cut/maximum flow in a graph. The proposed method is evaluated on scanned document images from the 18th century, showing an improvement of character recognition results compared to other restoration methods.

We present a new method for blind document bleed through removal based on separately restoring the recto and the verso side. The segmentation algorithm is based on separate Markov random fields (MRF) which results in a better adaptation of the prior to the content creation process (e.g. superimposing two pages), and the improvement of the estimation of the verso pixels through an estimation of the verso pixels covered by recto pixels. The labels of the initial recto and verso clusters are recognized without using any color or gray value information. The proposed method is evaluated empirically as well as through OCR improvement.

Content based image retrieval is the task of searching images from a database, which are visually similar to a given example image. We present methods for content based image retrieval based on texture similarity using interest points and Gabor features. Interest point detectors are used in computer vision to detect image points with special properties, which can be geometric (corners) or non-geometric (contrast etc.). Gabor functions and Gabor filters are regarded as excellent tools for feature extraction and texture segmentation. The article combines these methods and generates a textural description of images. Special emphasis is devoted to distance measures on texture descriptions. Experimental results of a query system are given.