Daniele De Martini

Benjamin Hardin, Efimia Panagiotaki, Daniele De Martini et al.

Although much is known about the physical danger of cycling situations, less is understood about the perceived danger of cycling. Furthermore, perception of danger may be filtered at a subconscious level and therefore difficult for one to self-report. To this end, these subconscious perceptions can be revealed through physiological metrics such as eye gaze. This paper explores the perceived safety of cycling in Oxford, United Kingdom and explores the ability of wearable eye tracking glasses to produce insights about the differences in perception under different environments and events. This paper finds that eye gaze patterns change between using bike lanes, car lanes and shared bus lanes, representing different cognitive challenges of each lane type. This paper presents that different intersections have significantly different eye gaze patterns which may have implications for cyclist stress. Finally, eye gaze patterns differ in the presence of events such as passes and pedestrians in the road compared to when cycling with no events. This paper draws conclusions on the benefits and limitations of using wearable eye trackers to estimate stress and cyclist workload.

Georgi Pramatarov, Daniele De Martini, Matthew Gadd et al. · oxford

This paper is about extremely robust and lightweight localisation using LiDAR point clouds based on instance segmentation and graph matching. We model 3D point clouds as fully-connected graphs of semantically identified components where each vertex corresponds to an object instance and encodes its shape. Optimal vertex association across graphs allows for full 6-Degree-of-Freedom (DoF) pose estimation and place recognition by measuring similarity. This representation is very concise, condensing the size of maps by a factor of 25 against the state-of-the-art, requiring only 3kB to represent a 1.4MB laser scan. We verify the efficacy of our system on the SemanticKITTI dataset, where we achieve a new state-of-the-art in place recognition, with an average of 88.4% recall at 100% precision where the next closest competitor follows with 64.9%. We also show accurate metric pose estimation performance - estimating 6-DoF pose with median errors of 10 cm and 0.33 deg.

Zhen Meng, Changyang She, Guodong Zhao et al.

The metaverse has the potential to revolutionize the next generation of the Internet by supporting highly interactive services with the help of Mixed Reality (MR) technologies; still, to provide a satisfactory experience for users, the synchronization between the physical world and its digital models is crucial. This work proposes a sampling, communication and prediction co-design framework to minimize the communication load subject to a constraint on tracking the Mean Squared Error (MSE) between a real-world device and its digital model in the metaverse. To optimize the sampling rate and the prediction horizon, we exploit expert knowledge and develop a constrained Deep Reinforcement Learning (DRL) algorithm, named Knowledge-assisted Constrained Twin-Delayed Deep Deterministic (KC-TD3) policy gradient algorithm. We validate our framework on a prototype composed of a real-world robotic arm and its digital model. Compared with existing approaches: (1) When the tracking error constraint is stringent (MSE=0.002 degrees), our policy degenerates into the policy in the sampling-communication co-design framework. (2) When the tracking error constraint is mild (MSE=0.007 degrees), our policy degenerates into the policy in the prediction-communication co-design framework. (3) Our framework achieves a better trade-off between the average MSE and the average communication load compared with a communication system without sampling and prediction. For example, the average communication load can be reduced up to 87% when the track error constraint is 0.002 degrees. (4) Our policy outperforms the benchmark with the static sampling rate and prediction horizon optimized by exhaustive search, in terms of the tail probability of the tracking error. Furthermore, with the assistance of expert knowledge, the proposed algorithm KC-TD3 achieves better convergence time, stability, and final policy performance.

Luigi Campanaro, Daniele De Martini, Siddhant Gangapurwala et al.

This paper proposes a simple strategy for sim-to-real in Deep-Reinforcement Learning (DRL) -- called Roll-Drop -- that uses dropout during simulation to account for observation noise during deployment without explicitly modelling its distribution for each state. DRL is a promising approach to control robots for highly dynamic and feedback-based manoeuvres, and accurate simulators are crucial to providing cheap and abundant data to learn the desired behaviour. Nevertheless, the simulated data are noiseless and generally show a distributional shift that challenges the deployment on real machines where sensor readings are affected by noise. The standard solution is modelling the latter and injecting it during training; while this requires a thorough system identification, Roll-Drop enhances the robustness to sensor noise by tuning only a single parameter. We demonstrate an 80% success rate when up to 25% noise is injected in the observations, with twice higher robustness than the baselines. We deploy the controller trained in simulation on a Unitree A1 platform and assess this improved robustness on the physical system.

Benjamin Ramtoula, Matthew Gadd, Paul Newman et al.

Selecting appropriate datasets is critical in modern computer vision. However, no general-purpose tools exist to evaluate the extent to which two datasets differ. For this, we propose representing images - and by extension datasets - using Distributions of Neuron Activations (DNAs). DNAs fit distributions, such as histograms or Gaussians, to activations of neurons in a pre-trained feature extractor through which we pass the image(s) to represent. This extractor is frozen for all datasets, and we rely on its generally expressive power in feature space. By comparing two DNAs, we can evaluate the extent to which two datasets differ with granular control over the comparison attributes of interest, providing the ability to customise the way distances are measured to suit the requirements of the task at hand. Furthermore, DNAs are compact, representing datasets of any size with less than 15 megabytes. We demonstrate the value of DNAs by evaluating their applicability on several tasks, including conditional dataset comparison, synthetic image evaluation, and transfer learning, and across diverse datasets, ranging from synthetic cat images to celebrity faces and urban driving scenes.

Pawit Kochakarn, Daniele De Martini, Daniel Omeiza et al.

This work explores scene graphs as a distilled representation of high-level information for autonomous driving, applied to future driver-action prediction. Given the scarcity and strong imbalance of data samples, we propose a self-supervision pipeline to infer representative and well-separated embeddings. Key aspects are interpretability and explainability; as such, we embed in our architecture attention mechanisms that can create spatial and temporal heatmaps on the scene graphs. We evaluate our system on the ROAD dataset against a fully-supervised approach, showing the superiority of our training regime.

Efimia Panagiotaki, Daniele De Martini, Lars Kunze

In this work, we propose a methodology for investigating the use of semantic attention to enhance the explainability of Graph Neural Network (GNN)-based models. Graph Deep Learning (GDL) has emerged as a promising field for tasks like scene interpretation, leveraging flexible graph structures to concisely describe complex features and relationships. As traditional explainability methods used in eXplainable AI (XAI) cannot be directly applied to such structures, graph-specific approaches are introduced. Attention has been previously employed to estimate the importance of input features in GDL, however, the fidelity of this method in generating accurate and consistent explanations has been questioned. To evaluate the validity of using attention weights as feature importance indicators, we introduce semantically-informed perturbations and correlate predicted attention weights with the accuracy of the model. Our work extends existing attention-based graph explainability methods by analysing the divergence in the attention distributions in relation to semantically sorted feature sets and the behaviour of a GNN model, efficiently estimating feature importance. We apply our methodology on a lidar pointcloud estimation model successfully identifying key semantic classes that contribute to enhanced performance, effectively generating reliable post-hoc semantic explanations.

Efimia Panagiotaki, Daniele De Martini, Georgi Pramatarov et al.

This paper proposes a Graph Neural Network(GNN)-based method for exploiting semantics and local geometry to guide the identification of reliable pointcloud registration candidates. Semantic and morphological features of the environment serve as key reference points for registration, enabling accurate lidar-based pose estimation. Our novel lightweight static graph structure informs our attention-based node aggregation network by identifying semantic-instance relationships, acting as an inductive bias to significantly reduce the computational burden of pointcloud registration. By connecting candidate nodes and exploiting cross-graph attention, we identify confidence scores for all potential registration correspondences and estimate the displacement between pointcloud scans. Our pipeline enables introspective analysis of the model's performance by correlating it with the individual contributions of local structures in the environment, providing valuable insights into the system's behaviour. We test our method on the KITTI odometry dataset, achieving competitive accuracy compared to benchmark methods and a higher track smoothness while relying on significantly fewer network parameters.

Valentina Musat, Daniele De Martini, Matthew Gadd et al.

In this paper we present a compositing image synthesis method that generates RGB canvases with well aligned segmentation maps and sparse depth maps, coupled with an in-painting network that transforms the RGB canvases into high quality RGB images and the sparse depth maps into pixel-wise dense depth maps. We benchmark our method in terms of structural alignment and image quality, showing an increase in mIoU over SOTA by 3.7 percentage points and a highly competitive FID. Furthermore, we analyse the quality of the generated data as training data for semantic segmentation and depth completion, and show that our approach is more suited for this purpose than other methods.

Matthew Gadd, Benjamin Ramtoula, Daniele De Martini et al.

Recalling the most relevant visual memories for localisation or understanding a priori the likely outcome of localisation effort against a particular visual memory is useful for efficient and robust visual navigation. Solutions to this problem should be divorced from performance appraisal against ground truth - as this is not available at run-time - and should ideally be based on generalisable environmental observations. For this, we propose applying the recently developed Visual DNA as a highly scalable tool for comparing datasets of images - in this work, sequences of map and live experiences. In the case of localisation, important dataset differences impacting performance are modes of appearance change, including weather, lighting, and season. Specifically, for any deep architecture which is used for place recognition by matching feature volumes at a particular layer, we use distribution measures to compare neuron-wise activation statistics between live images and multiple previously recorded past experiences, with a potentially large seasonal (winter/summer) or time of day (day/night) shift. We find that differences in these statistics correlate to performance when localising using a past experience with the same appearance gap. We validate our approach over the Nordland cross-season dataset as well as data from Oxford's University Parks with lighting and mild seasonal change, showing excellent ability of our system to rank actual localisation performance across candidate experiences.

Giorgia Modi, Davide Buoso, Giuseppe Averta et al.

Current approaches to 3D scene graph generation rely on dedicated depth sensors, such as LiDAR or RGB-D cameras, for metric 3D reconstruction. This limits deployment to specialized robotic platforms and excludes settings where only RGB cameras are available, such as fixed external infrastructure. Existing pipelines also typically operate on passively collected observation trajectories, rather than selecting viewpoints based on the partially built scene representation, and therefore fail to effectively exploit the semantic and spatial information encoded within the graph during exploration. This paper presents a fully visual framework for the active, incremental construction of 3D scene graphs from RGB input only, addressing both limitations. The proposed approach unifies perception and planning around a shared structured representation that captures object semantics, 3D geometry, relational context, and information from multiple viewpoints. Because the framework is hardware-agnostic and relies only on RGB observations, it can incorporate inputs from both onboard robot cameras and fixed external cameras within the same representation. Experiments on the Replica dataset show that the RGB-only pipeline achieves F1-score parity with baselines using ground-truth depth. Active exploration experiments on ReplicaCAD further show that semantic-driven viewpoint selection detects more than twice as many objects as a geometric frontier-based baseline under the same exploration budget. Finally, the external-camera setting demonstrates that complementary RGB views can effectively bootstrap the scene graph and improve contextual understanding at no additional exploration cost.

Giorgia Modi, Davide Buoso, Giuseppe Averta et al.

Commonly available prior information, such as BIM models, floor plans, and remote sensing images, can provide valuable geometric and semantic context for autonomous robotic systems. In this paper, we treat observations from fixed external RGB cameras as Common Prior Maps (CPMs): wide-field views of the environment that initialize a semantic and geometric scene prior before any robot motion begins. We present an RGB-only framework for active, incremental 3D scene graph (3DSG) generation that seamlessly fuses observations from both onboard robot cameras and fixed external cameras within a single hardware-agnostic pipeline. By relying solely on RGB observations processed by a feed-forward 3D reconstruction model, the system treats all cameras - onboard or external - identically, requiring no hardware modifications. A graph-based active semantic exploration framework then directly leverages the partial scene graph to guide the robot toward regions of high semantic uncertainty, progressively completing and refining the prior. Experiments demonstrate that bootstrapping the scene graph with even a single external camera increases initial object recall by up to +79%, and that the richer context of the prior significantly improves the efficiency of subsequent active exploration.

Benjamin Ramtoula, Pierre-Yves Lajoie, Paul Newman et al.

Foundation models (FMs) trained with different objectives and data learn diverse representations, making some more effective than others for specific downstream tasks. Existing adaptation strategies, such as parameter-efficient fine-tuning, focus on individual models and do not exploit the complementary strengths across models. Probing methods offer a promising alternative by extracting information from frozen models, but current techniques do not scale well with large feature sets and often rely on dataset-specific hyperparameter tuning. We propose Combined backBones (ComBo), a simple and scalable probing-based adapter that effectively integrates features from multiple models and layers. ComBo compresses activations from layers of one or more FMs into compact token-wise representations and processes them with a lightweight transformer for task-specific prediction. Crucially, ComBo does not require dataset-specific tuning or backpropagation through the backbone models. However, not all models are equally relevant for all tasks. To address this, we introduce a mechanism that leverages ComBo's joint multi-backbone probing to efficiently evaluate each backbone's task-relevance, enabling both practical model comparison and improved performance through selective adaptation. On the 19 tasks of the VTAB-1k benchmark, ComBo outperforms previous probing methods, matches or surpasses more expensive alternatives, such as distillation-based model merging, and enables efficient probing of tuned models. Our results demonstrate that ComBo offers a practical and general-purpose framework for combining diverse representations from multiple FMs.

Divya Thuremella, Yi Yang, Simon Wanna et al.

This work explores the application of ensemble modeling to the multidimensional regression problem of trajectory prediction for vehicles in urban environments. As newer and bigger state-of-the-art prediction models for autonomous driving continue to emerge, an important open challenge is the problem of how to combine the strengths of these big models without the need for costly re-training. We show how, perhaps surprisingly, combining state-of-the-art deep learning models out-of-the-box (without retraining or fine-tuning) with a simple confidence-weighted average method can enhance the overall prediction. Indeed, while combining trajectory prediction models is not straightforward, this simple approach enhances performance by 10% over the best prediction model, especially in the long-tailed metrics. We show that this performance improvement holds on both the NuScenes and Argoverse datasets, and that these improvements are made across the dataset distribution. The code for our work is open source.

David S. W. Williams, Daniele De Martini, Matthew Gadd et al.

Knowing when a trained segmentation model is encountering data that is different to its training data is important. Understanding and mitigating the effects of this play an important part in their application from a performance and assurance perspective - this being a safety concern in applications such as autonomous vehicles (AVs). This work presents a segmentation network that can detect errors caused by challenging test domains without any additional annotation in a single forward pass. As annotation costs limit the diversity of labelled datasets, we use easy-to-obtain, uncurated and unlabelled data to learn to perform uncertainty estimation by selectively enforcing consistency over data augmentation. To this end, a novel segmentation benchmark based on the SAX Dataset is used, which includes labelled test data spanning three autonomous-driving domains, ranging in appearance from dense urban to off-road. The proposed method, named Gamma-SSL, consistently outperforms uncertainty estimation and Out-of-Distribution (OoD) techniques on this difficult benchmark - by up to 10.7% in area under the receiver operating characteristic (ROC) curve and 19.2% in area under the precision-recall (PR) curve in the most challenging of the three scenarios.

Georgi Pramatarov, Matthew Gadd, Paul Newman et al.

This paper is about 3D pose estimation on LiDAR scans with extremely minimal storage requirements to enable scalable mapping and localisation. We achieve this by clustering all points of segmented scans into semantic objects and representing them only with their respective centroid and semantic class. In this way, each LiDAR scan is reduced to a compact collection of four-number vectors. This abstracts away important structural information from the scenes, which is crucial for traditional registration approaches. To mitigate this, we introduce an object-matching network based on self- and cross-correlation that captures geometric and semantic relationships between entities. The respective matches allow us to recover the relative transformation between scans through weighted Singular Value Decomposition (SVD) and RANdom SAmple Consensus (RANSAC). We demonstrate that such representation is sufficient for metric localisation by registering point clouds taken under different viewpoints on the KITTI dataset, and at different periods of time localising between KITTI and KITTI-360. We achieve accurate metric estimates comparable with state-of-the-art methods with almost half the representation size, specifically 1.33 kB on average.

Davide Buoso, Luke Robinson, Giuseppe Averta et al.

This study explores the potential of off-the-shelf Vision-Language Models (VLMs) for high-level robot planning in the context of autonomous navigation. Indeed, while most of existing learning-based approaches for path planning require extensive task-specific training/fine-tuning, we demonstrate how such training can be avoided for most practical cases. To do this, we introduce Select2Plan (S2P), a novel training-free framework for high-level robot planning which completely eliminates the need for fine-tuning or specialised training. By leveraging structured Visual Question-Answering (VQA) and In-Context Learning (ICL), our approach drastically reduces the need for data collection, requiring a fraction of the task-specific data typically used by trained models, or even relying only on online data. Our method facilitates the effective use of a generally trained VLM in a flexible and cost-efficient way, and does not require additional sensing except for a simple monocular camera. We demonstrate its adaptability across various scene types, context sources, and sensing setups. We evaluate our approach in two distinct scenarios: traditional First-Person View (FPV) and infrastructure-driven Third-Person View (TPV) navigation, demonstrating the flexibility and simplicity of our method. Our technique significantly enhances the navigational capabilities of a baseline VLM of approximately 50% in TPV scenario, and is comparable to trained models in the FPV one, with as few as 20 demonstrations.

Efimia Panagiotaki, Daniele De Martini, Lars Kunze et al.

This study explores the intersection of neural networks and classical robotics algorithms through the Neural Algorithmic Reasoning (NAR) blueprint, enabling the training of neural networks to reason like classical robotics algorithms by learning to execute them. Algorithms are integral to robotics and safety-critical applications due to their predictable and consistent performance through logical and mathematical principles. In contrast, while neural networks are highly adaptable, handling complex, high-dimensional data and generalising across tasks, they often lack interpretability and transparency in their internal computations. To bridge the two, we propose a novel Graph Neural Network (GNN)-based framework, NAR-*ICP, that learns the intermediate computations of classical ICP-based registration algorithms, extending the CLRS Benchmark. We evaluate our approach across real-world and synthetic datasets, demonstrating its flexibility in handling complex inputs, and its potential to be used within larger learning pipelines. Our method achieves superior performance compared to the baselines, even surpassing the algorithms it was trained on, further demonstrating its ability to generalise beyond the capabilities of traditional algorithms.

Samuel Sze, Daniele De Martini, Lars Kunze

Developing 3D semantic occupancy prediction models often relies on dense 3D annotations for supervised learning, a process that is both labor and resource-intensive, underscoring the need for label-efficient or even label-free approaches. To address this, we introduce MinkOcc, a multi-modal 3D semantic occupancy prediction framework for cameras and LiDARs that proposes a two-step semi-supervised training procedure. Here, a small dataset of explicitly 3D annotations warm-starts the training process; then, the supervision is continued by simpler-to-annotate accumulated LiDAR sweeps and images -- semantically labelled through vision foundational models. MinkOcc effectively utilizes these sensor-rich supervisory cues and reduces reliance on manual labeling by 90\% while maintaining competitive accuracy. In addition, the proposed model incorporates information from LiDAR and camera data through early fusion and leverages sparse convolution networks for real-time prediction. With its efficiency in both supervision and computation, we aim to extend MinkOcc beyond curated datasets, enabling broader real-world deployment of 3D semantic occupancy prediction in autonomous driving.

Yufeng Diao, Yichi Zhang, Daniele De Martini et al.

This paper proposes a task-oriented co-design framework that integrates communication, computing, and control to address the key challenges of bandwidth limitations, noise interference, and latency in mission-critical industrial Cyber-Physical Systems (CPS). To improve communication efficiency and robustness, we design a task-oriented Joint Source-Channel Coding (JSCC) using Information Bottleneck (IB) to enhance data transmission efficiency by prioritizing task-specific information. To mitigate the perceived End-to-End (E2E) delays, we develop a Delay-Aware Trajectory-Guided Control Prediction (DTCP) strategy that integrates trajectory planning with control prediction, predicting commands based on E2E delay. Moreover, the DTCP is co-designed with task-oriented JSCC, focusing on transmitting task-specific information for timely and reliable autonomous driving. Experimental results in the CARLA simulator demonstrate that, under an E2E delay of 1 second (20 time slots), the proposed framework achieves a driving score of 48.12, which is 31.59 points higher than using Better Portable Graphics (BPG) while reducing bandwidth usage by 99.19%.

David S. W. Williams, Matthew Gadd, Paul Newman et al.

This work proposes a semantic segmentation network that produces high-quality uncertainty estimates in a single forward pass. We exploit general representations from foundation models and unlabelled datasets through a Masked Image Modeling (MIM) approach, which is robust to augmentation hyper-parameters and simpler than previous techniques. For neural networks used in safety-critical applications, bias in the training data can lead to errors; therefore it is crucial to understand a network's limitations at run time and act accordingly. To this end, we test our proposed method on a number of test domains including the SAX Segmentation benchmark, which includes labelled test data from dense urban, rural and off-road driving domains. The proposed method consistently outperforms uncertainty estimation and Out-of-Distribution (OoD) techniques on this difficult benchmark.

Jianhao Yuan, Fabio Pizzati, Francesco Pinto et al.

Intuitive physics understanding in video diffusion models plays an essential role in building general-purpose physically plausible world simulators, yet accurately evaluating such capacity remains a challenging task due to the difficulty in disentangling physics correctness from visual appearance in generation. To the end, we introduce LikePhys, a training-free method that evaluates intuitive physics in video diffusion models by distinguishing physically valid and impossible videos using the denoising objective as an ELBO-based likelihood surrogate on a curated dataset of valid-invalid pairs. By testing on our constructed benchmark of twelve scenarios spanning over four physics domains, we show that our evaluation metric, Plausibility Preference Error (PPE), demonstrates strong alignment with human preference, outperforming state-of-the-art evaluator baselines. We then systematically benchmark intuitive physics understanding in current video diffusion models. Our study further analyses how model design and inference settings affect intuitive physics understanding and highlights domain-specific capacity variations across physical laws. Empirical results show that, despite current models struggling with complex and chaotic dynamics, there is a clear trend of improvement in physics understanding as model capacity and inference settings scale.

Manshika Charvi Bissessur, Efimia Panagiotaki, Daniele De Martini

This work investigates how semantics influence localisation performance and robustness in a learned self-supervised, contrastive semantic localisation framework. After training a localisation network on both original and perturbed maps, we conduct a thorough post-hoc introspection analysis to probe whether the model filters environmental noise and prioritises distinctive landmarks over routine clutter. We validate various interpretability methods and present a comparative reliability analysis. Integrated gradients and Attention Weights consistently emerge as the most reliable probes of learned behaviour. A semantic class ablation further reveals an implicit weighting in which frequent objects are often down-weighted. Overall, the results indicate that the model learns noise-robust, semantically salient relations about place definition, thereby enabling explainable registration under challenging visual and structural variations.

Sitian Shen, Georgi Pramatarov, Yifu Tao et al.

This paper proposes Neural-MMGS, a novel neural 3DGS framework for multimodal large-scale scene reconstruction that fuses multiple sensing modalities in a per-gaussian compact, learnable embedding. While recent works focusing on large-scale scene reconstruction have incorporated LiDAR data to provide more accurate geometric constraints, we argue that LiDAR's rich physical properties remain underexplored. Similarly, semantic information has been used for object retrieval, but could provide valuable high-level context for scene reconstruction. Traditional approaches append these properties to Gaussians as separate parameters, increasing memory usage and limiting information exchange across modalities. Instead, our approach fuses all modalities -- image, LiDAR, and semantics -- into a compact, learnable embedding that implicitly encodes optical, physical, and semantic features in each Gaussian. We then train lightweight neural decoders to map these embeddings to Gaussian parameters, enabling the reconstruction of each sensing modality with lower memory overhead and improved scalability. We evaluate Neural-MMGS on the Oxford Spires and KITTI-360 datasets. On Oxford Spires, we achieve higher-quality reconstructions, while on KITTI-360, our method reaches competitive results with less storage consumption compared with current approaches in LiDAR-based novel-view synthesis.

Efimia Panagiotaki, Georgi Pramatarov, Lars Kunze et al.

Testing and validating Autonomous Vehicle (AV) performance in safety-critical and diverse scenarios is crucial before real-world deployment. However, manually creating such scenarios in simulation remains a significant and time-consuming challenge. This work introduces a novel method that generates dynamic temporal scene graphs corresponding to diverse traffic scenarios, on-demand, tailored to user-defined preferences, such as AV actions, sets of dynamic agents, and criticality levels. A temporal Graph Neural Network (GNN) model learns to predict relationships between ego-vehicle, agents, and static structures, guided by real-world spatiotemporal interaction patterns and constrained by an ontology that restricts predictions to semantically valid links. Our model consistently outperforms the baselines in accurately generating links corresponding to the requested scenarios. We render the predicted scenarios in simulation to further demonstrate their effectiveness as testing environments for AV agents.

Benjamin Ramtoula, Daniele De Martini, Matthew Gadd et al.

This paper adapts a general dataset representation technique to produce robust Visual Place Recognition (VPR) descriptors, crucial to enable real-world mobile robot localisation. Two parallel lines of work on VPR have shown, on one side, that general-purpose off-the-shelf feature representations can provide robustness to domain shifts, and, on the other, that fused information from sequences of images improves performance. In our recent work on measuring domain gaps between image datasets, we proposed a Visual Distribution of Neuron Activations (VDNA) representation to represent datasets of images. This representation can naturally handle image sequences and provides a general and granular feature representation derived from a general-purpose model. Moreover, our representation is based on tracking neuron activation values over the list of images to represent and is not limited to a particular neural network layer, therefore having access to high- and low-level concepts. This work shows how VDNAs can be used for VPR by learning a very lightweight and simple encoder to generate task-specific descriptors. Our experiments show that our representation can allow for better robustness than current solutions to serious domain shifts away from the training data distribution, such as to indoor environments and aerial imagery.

Vít Růžička, Anna Vaughan, Daniele De Martini et al.

In this paper, we introduce RaVAEn, a lightweight, unsupervised approach for change detection in satellite data based on Variational Auto-Encoders (VAEs) with the specific purpose of on-board deployment. Applications such as disaster management enormously benefit from the rapid availability of satellite observations. Traditionally, data analysis is performed on the ground after all data is transferred - downlinked - to a ground station. Constraint on the downlink capabilities therefore affects any downstream application. In contrast, RaVAEn pre-processes the sampled data directly on the satellite and flags changed areas to prioritise for downlink, shortening the response time. We verified the efficacy of our system on a dataset composed of time series of catastrophic events - which we plan to release alongside this publication - demonstrating that RaVAEn outperforms pixel-wise baselines. Finally we tested our approach on resource-limited hardware for assessing computational and memory limitations.

Matthew Gadd, Daniele De Martini, Paul Newman

We learn, in an unsupervised way, an embedding from sequences of radar images that is suitable for solving the place recognition problem with complex radar data. Our method is based on invariant instance feature learning but is tailored for the task of re-localisation by exploiting for data augmentation the temporal successivity of data as collected by a mobile platform moving through the scene smoothly. We experiment across two prominent urban radar datasets totalling over 400 km of driving and show that we achieve a new radar place recognition state-of-the-art. Specifically, the proposed system proves correct for 98.38% of the queries that it is presented with over a challenging re-localisation sequence, using only the single nearest neighbour in the learned metric space. We also find that our learned model shows better understanding of out-of-lane loop closures at arbitrary orientation than non-learned radar scan descriptors.

Tarlan Suleymanov, Matthew Gadd, Daniele De Martini et al.

In this paper we present the Oxford Road Boundaries Dataset, designed for training and testing machine-learning-based road-boundary detection and inference approaches. We have hand-annotated two of the 10 km-long forays from the Oxford Robotcar Dataset and generated from other forays several thousand further examples with semi-annotated road-boundary masks. To boost the number of training samples in this way, we used a vision-based localiser to project labels from the annotated datasets to other traversals at different times and weather conditions. As a result, we release 62605 labelled samples, of which 47639 samples are curated. Each of these samples contains both raw and classified masks for left and right lenses. Our data contains images from a diverse set of scenarios such as straight roads, parked cars, junctions, etc. Files for download and tools for manipulating the labelled data are available at: oxford-robotics-institute.github.io/road-boundaries-dataset

Matthew Gadd, Daniele De Martini, Paul Newman

We learn, in an unsupervised way, an embedding from sequences of radar images that is suitable for solving place recognition problem using complex radar data. We experiment on 280 km of data and show performance exceeding state-of-the-art supervised approaches, localising correctly 98.38% of the time when using just the nearest database candidate.

David Williams, Matthew Gadd, Daniele De Martini et al.

In this work, we train a network to simultaneously perform segmentation and pixel-wise Out-of-Distribution (OoD) detection, such that the segmentation of unknown regions of scenes can be rejected. This is made possible by leveraging an OoD dataset with a novel contrastive objective and data augmentation scheme. By combining data including unknown classes in the training data, a more robust feature representation can be learned with known classes represented distinctly from those unknown. When presented with unknown classes or conditions, many current approaches for segmentation frequently exhibit high confidence in their inaccurate segmentations and cannot be trusted in many operational environments. We validate our system on a real-world dataset of unusual driving scenes, and show that by selectively segmenting scenes based on what is predicted as OoD, we can increase the segmentation accuracy by an IoU of 0.2 with respect to alternative techniques.

Luigi Campanaro, Siddhant Gangapurwala, Daniele De Martini et al.

Central Pattern Generators (CPGs) have several properties desirable for locomotion: they generate smooth trajectories, are robust to perturbations and are simple to implement. Although conceptually promising, we argue that the full potential of CPGs has so far been limited by insufficient sensory-feedback information. This paper proposes a new methodology that allows tuning CPG controllers through gradient-based optimization in a Reinforcement Learning (RL) setting. To the best of our knowledge, this is the first time CPGs have been trained in conjunction with a MultilayerPerceptron (MLP) network in a Deep-RL context. In particular, we show how CPGs can directly be integrated as the Actor in an Actor-Critic formulation. Additionally, we demonstrate how this change permits us to integrate highly non-linear feedback directly from sensory perception to reshape the oscillators' dynamics. Our results on a locomotion task using a single-leg hopper demonstrate that explicitly using the CPG as the Actor rather than as part of the environment results in a significant increase in the reward gained over time (6x more) compared with previous approaches. Furthermore, we show that our method without feedback reproduces results similar to prior work with feedback. Finally, we demonstrate how our closed-loop CPG progressively improves the hopping behaviour for longer training epochs relying only on basic reward functions.

Christian Schroeder de Witt, Catherine Tong, Valentina Zantedeschi et al.

Extreme precipitation events, such as violent rainfall and hail storms, routinely ravage economies and livelihoods around the developing world. Climate change further aggravates this issue. Data-driven deep learning approaches could widen the access to accurate multi-day forecasts, to mitigate against such events. However, there is currently no benchmark dataset dedicated to the study of global precipitation forecasts. In this paper, we introduce \textbf{RainBench}, a new multi-modal benchmark dataset for data-driven precipitation forecasting. It includes simulated satellite data, a selection of relevant meteorological data from the ERA5 reanalysis product, and IMERG precipitation data. We also release \textbf{PyRain}, a library to process large precipitation datasets efficiently. We present an extensive analysis of our novel dataset and establish baseline results for two benchmark medium-range precipitation forecasting tasks. Finally, we discuss existing data-driven weather forecasting methodologies and suggest future research avenues.

Tim Y. Tang, Daniele De Martini, Shangzhe Wu et al.

Publicly available satellite imagery can be an ubiquitous, cheap, and powerful tool for vehicle localisation when a prior sensor map is unavailable. However, satellite images are not directly comparable to data from ground range sensors because of their starkly different modalities. We present a learned metric localisation method that not only handles the modality difference, but is cheap to train, learning in a self-supervised fashion without metrically accurate ground truth. By evaluating across multiple real-world datasets, we demonstrate the robustness and versatility of our method for various sensor configurations. We pay particular attention to the use of millimetre wave radar, which, owing to its complex interaction with the scene and its immunity to weather and lighting, makes for a compelling and valuable use case.

David Williams, Daniele De Martini, Matthew Gadd et al.

Reliable outdoor deployment of mobile robots requires the robust identification of permissible driving routes in a given environment. The performance of LiDAR and vision-based perception systems deteriorates significantly if certain environmental factors are present e.g. rain, fog, darkness. Perception systems based on FMCW scanning radar maintain full performance regardless of environmental conditions and with a longer range than alternative sensors. Learning to segment a radar scan based on driveability in a fully supervised manner is not feasible as labelling each radar scan on a bin-by-bin basis is both difficult and time-consuming to do by hand. We therefore weakly supervise the training of the radar-based classifier through an audio-based classifier that is able to predict the terrain type underneath the robot. By combining odometry, GPS and the terrain labels from the audio classifier, we are able to construct a terrain labelled trajectory of the robot in the environment which is then used to label the radar scans. Using a curriculum learning procedure, we then train a radar segmentation network to generalise beyond the initial labelling and to detect all permissible driving routes in the environment.

Matthew Gadd, Daniele De Martini, Letizia Marchegiani et al.

This paper discusses ongoing work in demonstrating research in mobile autonomy in challenging driving scenarios. In our approach, we address fundamental technical issues to overcome critical barriers to assurance and regulation for large-scale deployments of autonomous systems. To this end, we present how we build robots that (1) can robustly sense and interpret their environment using traditional as well as unconventional sensors; (2) can assess their own capabilities; and (3), vitally in the purpose of assurance and trust, can provide causal explanations of their interpretations and assessments. As it is essential that robots are safe and trusted, we design, develop, and demonstrate fundamental technologies in real-world applications to overcome critical barriers which impede the current deployment of robots in economically and socially important areas. Finally, we describe ongoing work in the collection of an unusual, rare, and highly valuable dataset.

Prannay Kaul, Daniele De Martini, Matthew Gadd et al.

This paper presents an efficient annotation procedure and an application thereof to end-to-end, rich semantic segmentation of the sensed environment using FMCW scanning radar. We advocate radar over the traditional sensors used for this task as it operates at longer ranges and is substantially more robust to adverse weather and illumination conditions. We avoid laborious manual labelling by exploiting the largest radar-focused urban autonomy dataset collected to date, correlating radar scans with RGB cameras and LiDAR sensors, for which semantic segmentation is an already consolidated procedure. The training procedure leverages a state-of-the-art natural image segmentation system which is publicly available and as such, in contrast to previous approaches, allows for the production of copious labels for the radar stream by incorporating four camera and two LiDAR streams. Additionally, the losses are computed taking into account labels to the radar sensor horizon by accumulating LiDAR returns along a pose-chain ahead and behind of the current vehicle position. Finally, we present the network with multi-channel radar scan inputs in order to deal with ephemeral and dynamic scene objects.

Matthew Gadd, Daniele De Martini, Paul Newman

This paper details an application which yields significant improvements to the adeptness of place recognition with Frequency-Modulated Continuous-Wave radar - a commercially promising sensor poised for exploitation in mobile autonomy. We show how a rotationally-invariant metric embedding for radar scans can be integrated into sequence-based trajectory matching systems typically applied to videos taken by visual sensors. Due to the complete horizontal field of view inherent to the radar scan formation process, we show how this off-the-shelf sequence-based trajectory matching system can be manipulated to detect place matches when the vehicle is travelling down a previously visited stretch of road in the opposite direction. We demonstrate the efficacy of the approach on 26 km of challenging urban driving taken from the largest radar-focused urban autonomy dataset released to date -- showing a boost of 30% in recall at high levels of precision over a nearest neighbour approach.

Ştefan Săftescu, Matthew Gadd, Daniele De Martini et al.

This paper presents a system for robust, large-scale topological localisation using Frequency-Modulated Continuous-Wave (FMCW) scanning radar. We learn a metric space for embedding polar radar scans using CNN and NetVLAD architectures traditionally applied to the visual domain. However, we tailor the feature extraction for more suitability to the polar nature of radar scan formation using cylindrical convolutions, anti-aliasing blurring, and azimuth-wise max-pooling; all in order to bolster the rotational invariance. The enforced metric space is then used to encode a reference trajectory, serving as a map, which is queried for nearest neighbours (NNs) for recognition of places at run-time. We demonstrate the performance of our topological localisation system over the course of many repeat forays using the largest radar-focused mobile autonomy dataset released to date, totalling 280 km of urban driving, a small portion of which we also use to learn the weights of the modified architecture. As this work represents a novel application for FMCW radar, we analyse the utility of the proposed method via a comprehensive set of metrics which provide insight into the efficacy when used in a realistic system, showing improved performance over the root architecture even in the face of random rotational perturbation.

Tim Y. Tang, Daniele De Martini, Dan Barnes et al.

This paper is about localising a vehicle in an overhead image using FMCW radar mounted on a ground vehicle. FMCW radar offers extraordinary promise and efficacy for vehicle localisation. It is impervious to all weather types and lighting conditions. However the complexity of the interactions between millimetre radar wave and the physical environment makes it a challenging domain. Infrastructure-free large-scale radar-based localisation is in its infancy. Typically here a map is built and suitable techniques, compatible with the nature of sensor, are brought to bear. In this work we eschew the need for a radar-based map; instead we simply use an overhead image -- a resource readily available everywhere. This paper introduces a method that not only naturally deals with the complexity of the signal type but does so in the context of cross modal processing.

Mirto Musci, Daniele De Martini, Nicola Blago et al.

Unintentional falls can cause severe injuries and even death, especially if no immediate assistance is given. The aim of Fall Detection Systems (FDSs) is to detect an occurring fall. This information can be used to trigger the necessary assistance in case of injury. This can be done by using either ambient-based sensors, e.g. cameras, or wearable devices. The aim of this work is to study the technical aspects of FDSs based on wearable devices and artificial intelligence techniques, in particular Deep Learning (DL), to implement an effective algorithm for on-line fall detection. The proposed classifier is based on a Recurrent Neural Network (RNN) model with underlying Long Short-Term Memory (LSTM) blocks. The method is tested on the publicly available SisFall dataset, with extended annotation, and compared with the results obtained by the SisFall authors.