Tom Drummond

Xuelian Cheng, Huan Xiong, Deng-Ping Fan et al. · ibm-research

We propose a new video camouflaged object detection (VCOD) framework that can exploit both short-term dynamics and long-term temporal consistency to detect camouflaged objects from video frames. An essential property of camouflaged objects is that they usually exhibit patterns similar to the background and thus make them hard to identify from still images. Therefore, effectively handling temporal dynamics in videos becomes the key for the VCOD task as the camouflaged objects will be noticeable when they move. However, current VCOD methods often leverage homography or optical flows to represent motions, where the detection error may accumulate from both the motion estimation error and the segmentation error. On the other hand, our method unifies motion estimation and object segmentation within a single optimization framework. Specifically, we build a dense correlation volume to implicitly capture motions between neighbouring frames and utilize the final segmentation supervision to optimize the implicit motion estimation and segmentation jointly. Furthermore, to enforce temporal consistency within a video sequence, we jointly utilize a spatio-temporal transformer to refine the short-term predictions. Extensive experiments on VCOD benchmarks demonstrate the architectural effectiveness of our approach. We also provide a large-scale VCOD dataset named MoCA-Mask with pixel-level handcrafted ground-truth masks and construct a comprehensive VCOD benchmark with previous methods to facilitate research in this direction. Dataset Link: https://xueliancheng.github.io/SLT-Net-project.

Xuelian Cheng, Yiran Zhong, Mehrtash Harandi et al. · ibm-research

The self-attention mechanism, successfully employed with the transformer structure is shown promise in many computer vision tasks including image recognition, and object detection. Despite the surge, the use of the transformer for the problem of stereo matching remains relatively unexplored. In this paper, we comprehensively investigate the use of the transformer for the problem of stereo matching, especially for laparoscopic videos, and propose a new hybrid deep stereo matching framework (HybridStereoNet) that combines the best of the CNN and the transformer in a unified design. To be specific, we investigate several ways to introduce transformers to volumetric stereo matching pipelines by analyzing the loss landscape of the designs and in-domain/cross-domain accuracy. Our analysis suggests that employing transformers for feature representation learning, while using CNNs for cost aggregation will lead to faster convergence, higher accuracy and better generalization than other options. Our extensive experiments on Sceneflow, SCARED2019 and dVPN datasets demonstrate the superior performance of our HybridStereoNet.

Nicholas Rosa, Tom Drummond, Mehrtash Harandi

Naively trained AI models can be heavily biased. This can be particularly problematic when the biases involve legally or morally protected attributes such as ethnic background, age or gender. Existing solutions to this problem come at the cost of extra computation, unstable adversarial optimisation or have losses on the feature space structure that are disconnected from fairness measures and only loosely generalise to fairness. In this work we propose a differentiable approximation of the variance of demographics, a metric that can be used to measure the bias, or unfairness, in an AI model. Our approximation can be optimised alongside the regular training objective which eliminates the need for any extra models during training and directly improves the fairness of the regularised models. We demonstrate that our approach improves the fairness of AI models in varied task and dataset scenarios, whilst still maintaining a high level of classification accuracy. Code is available at https://bitbucket.org/nelliottrosa/base_fairness.

Markus Hiller, Rongkai Ma, Mehrtash Harandi et al.

Single image-level annotations only correctly describe an often small subset of an image's content, particularly when complex real-world scenes are depicted. While this might be acceptable in many classification scenarios, it poses a significant challenge for applications where the set of classes differs significantly between training and test time. In this paper, we take a closer look at the implications in the context of $\textit{few-shot learning}$. Splitting the input samples into patches and encoding these via the help of Vision Transformers allows us to establish semantic correspondences between local regions across images and independent of their respective class. The most informative patch embeddings for the task at hand are then determined as a function of the support set via online optimization at inference time, additionally providing visual interpretability of `$\textit{what matters most}$' in the image. We build on recent advances in unsupervised training of networks via masked image modelling to overcome the lack of fine-grained labels and learn the more general statistical structure of the data while avoiding negative image-level annotation influence, $\textit{aka}$ supervision collapse. Experimental results show the competitiveness of our approach, achieving new state-of-the-art results on four popular few-shot classification benchmarks for $5$-shot and $1$-shot scenarios.

Tianyu Zhu, Bryce Ferenczi, Pulak Purkait et al.

Rotated bounding boxes drastically reduce output ambiguity of elongated objects, making it superior to axis-aligned bounding boxes. Despite the effectiveness, rotated detectors are not widely employed. Annotating rotated bounding boxes is such a laborious process that they are not provided in many detection datasets where axis-aligned annotations are used instead. In this paper, we propose a framework that allows the model to predict precise rotated boxes only requiring cheaper axis-aligned annotation of the target dataset 1. To achieve this, we leverage the fact that neural networks are capable of learning richer representation of the target domain than what is utilized by the task. The under-utilized representation can be exploited to address a more detailed task. Our framework combines task knowledge of an out-of-domain source dataset with stronger annotation and domain knowledge of the target dataset with weaker annotation. A novel assignment process and projection loss are used to enable the co-training on the source and target datasets. As a result, the model is able to solve the more detailed task in the target domain, without additional computation overhead during inference. We extensively evaluate the method on various target datasets including fresh-produce dataset, HRSC2016 and SSDD. Results show that the proposed method consistently performs on par with the fully supervised approach.

Markus Hiller, Mehrtash Harandi, Tom Drummond

Inspired by the concept of preconditioning, we propose a novel method to increase adaptation speed for gradient-based meta-learning methods without incurring extra parameters. We demonstrate that recasting the optimization problem to a non-linear least-squares formulation provides a principled way to actively enforce a $\textit{well-conditioned}$ parameter space for meta-learning models based on the concepts of the condition number and local curvature. Our comprehensive evaluations show that the proposed method significantly outperforms its unconstrained counterpart especially during initial adaptation steps, while achieving comparable or better overall results on several few-shot classification tasks -- creating the possibility of dynamically choosing the number of adaptation steps at inference time.

Luke Ditria, Tom Drummond

Generative models have seen an explosion in popularity with the release of huge generative Diffusion models like Midjourney and Stable Diffusion to the public. Because of this new ease of access, questions surrounding the automated collection of data and issues regarding content ownership have started to build. In this paper we present new work which aims to provide ways of protecting content when shared to the public. We show that a generative Diffusion model trained on data that has been imperceptibly watermarked will generate new images with these watermarks present. We further show that if a given watermark is correlated with a certain feature of the training data, the generated images will also have this correlation. Using statistical tests we show that we are able to determine whether a model has been trained on marked data, and what data was marked. As a result our system offers a solution to protect intellectual property when sharing content online.

Luke Ditria, Tom Drummond

Predicting high dimensional video sequences is a curiously difficult problem. The number of possible futures for a given video sequence grows exponentially over time due to uncertainty. This is especially evident when trying to predict complicated natural video scenes from a limited snapshot of the world. The inherent uncertainty accumulates the further into the future you predict making long-term prediction very difficult. In this work we introduce a number of improvements to existing work that aid in creating Robust Video Predictors (RoViPs). We show that with a combination of deep Perceptual and uncertainty-based reconstruction losses we are able to create high quality short-term predictions. Attention-based skip connections are utilised to allow for long range spatial movement of input features to further improve performance. Finally, we show that by simply making the predictor robust to its own prediction errors, it is possible to produce very long, realistic natural video sequences using an iterated single-step prediction task.

Fan Jiang, Qiongkai Xu, Tom Drummond et al.

Neural 'dense' retrieval models are state of the art for many datasets, however these models often exhibit limited domain transfer ability. Existing approaches to adaptation are unwieldy, such as requiring explicit supervision, complex model architectures, or massive external models. We present $\texttt{ABEL}$, a simple but effective unsupervised method to enhance passage retrieval in zero-shot settings. Our technique follows a straightforward loop: a dense retriever learns from supervision signals provided by a reranker, and subsequently, the reranker is updated based on feedback from the improved retriever. By iterating this loop, the two components mutually enhance one another's performance. Experimental results demonstrate that our unsupervised $\texttt{ABEL}$ model outperforms both leading supervised and unsupervised retrievers on the BEIR benchmark. Meanwhile, it exhibits strong adaptation abilities to tasks and domains that were unseen during training. By either fine-tuning $\texttt{ABEL}$ on labelled data or integrating it with existing supervised dense retrievers, we achieve state-of-the-art results.\footnote{Source code is available at \url{https://github.com/Fantabulous-J/BootSwitch}.}

Yunyan Xing, Benjamin J. Meyer, Mehrtash Harandi et al.

Content-based medical image retrieval is an important diagnostic tool that improves the explainability of computer-aided diagnosis systems and provides decision making support to healthcare professionals. Medical imaging data, such as radiology images, are often multimorbidity; a single sample may have more than one pathology present. As such, image retrieval systems for the medical domain must be designed for the multi-label scenario. In this paper, we propose a novel multi-label metric learning method that can be used for both classification and content-based image retrieval. In this way, our model is able to support diagnosis by predicting the presence of diseases and provide evidence for these predictions by returning samples with similar pathological content to the user. In practice, the retrieved images may also be accompanied by pathology reports, further assisting in the diagnostic process. Our method leverages proxy feature vectors, enabling the efficient learning of a robust feature space in which the distance between feature vectors can be used as a measure of the similarity of those samples. Unlike existing proxy-based methods, training samples are able to assign to multiple proxies that span multiple class labels. This multi-label proxy assignment results in a feature space that encodes the complex relationships between diseases present in medical imaging data. Our method outperforms state-of-the-art image retrieval systems and a set of baseline approaches. We demonstrate the efficacy of our approach to both classification and content-based image retrieval on two multimorbidity radiology datasets.

Fan Jiang, Tom Drummond, Trevor Cohn

Although existing neural retrieval models reveal promising results when training data is abundant and the performance keeps improving as training data increases, collecting high-quality annotated data is prohibitively costly. To this end, we introduce a novel noisy self-training framework combined with synthetic queries, showing that neural retrievers can be improved in a self-evolution manner with no reliance on any external models. Experimental results show that our method improves consistently over existing methods on both general-domain (e.g., MS-MARCO) and out-of-domain (i.e., BEIR) retrieval benchmarks. Extra analysis on low-resource settings reveals that our method is data efficient and outperforms competitive baselines, with as little as 30% of labelled training data. Further extending the framework for reranker training demonstrates that the proposed method is general and yields additional gains on tasks of diverse domains.\footnote{Source code is available at \url{https://github.com/Fantabulous-J/Self-Training-DPR}}

Zhenkai Zhang, Krista A. Ehinger, Tom Drummond

This paper introduces two key contributions aimed at improving the speed and quality of images generated through inverse diffusion processes. The first contribution involves reparameterizing the diffusion process in terms of the angle on a quarter-circular arc between the image and noise, specifically setting the conventional $\displaystyle \sqrt{\barα}=\cos(η)$. This reparameterization eliminates two singularities and allows for the expression of diffusion evolution as a well-behaved ordinary differential equation (ODE). In turn, this allows higher order ODE solvers such as Runge-Kutta methods to be used effectively. The second contribution is to directly estimate both the image ($\mathbf{x}_0$) and noise ($\mathbfε$) using our network, which enables more stable calculations of the update step in the inverse diffusion steps, as accurate estimation of both the image and noise are crucial at different stages of the process. Together with these changes, our model achieves faster generation, with the ability to converge on high-quality images more quickly, and higher quality of the generated images, as measured by metrics such as Frechet Inception Distance (FID), spatial Frechet Inception Distance (sFID), precision, and recall.

Bryce Ferenczi, Rhys Newbury, Michael Burke et al.

Creation and storage of datasets are often overlooked input costs in machine learning, as many datasets are simple image label pairs or plain text. However, datasets with more complex structures, such as those from the real time strategy game StarCraft II, require more deliberate thought and strategy to reduce cost of ownership. We introduce a serialization framework for StarCraft II that reduces the cost of dataset creation and storage, as well as improving usage ergonomics. We benchmark against the most comparable existing dataset from \textit{AlphaStar-Unplugged} and highlight the benefit of our framework in terms of both the cost of creation and storage. We use our dataset to train deep learning models that exceed the performance of comparable models trained on other datasets. The dataset conversion and usage framework introduced is open source and can be used as a framework for datasets with similar characteristics such as digital twin simulations. Pre-converted StarCraft II tournament data is also available online.

Xuelian Cheng, Yiran Zhong, Mehrtash Harandi et al.

To reduce the human efforts in neural network design, Neural Architecture Search (NAS) has been applied with remarkable success to various high-level vision tasks such as classification and semantic segmentation. The underlying idea for the NAS algorithm is straightforward, namely, to enable the network the ability to choose among a set of operations (e.g., convolution with different filter sizes), one is able to find an optimal architecture that is better adapted to the problem at hand. However, so far the success of NAS has not been enjoyed by low-level geometric vision tasks such as stereo matching. This is partly due to the fact that state-of-the-art deep stereo matching networks, designed by humans, are already sheer in size. Directly applying the NAS to such massive structures is computationally prohibitive based on the currently available mainstream computing resources. In this paper, we propose the first end-to-end hierarchical NAS framework for deep stereo matching by incorporating task-specific human knowledge into the neural architecture search framework. Specifically, following the gold standard pipeline for deep stereo matching (i.e., feature extraction -- feature volume construction and dense matching), we optimize the architectures of the entire pipeline jointly. Extensive experiments show that our searched network outperforms all state-of-the-art deep stereo matching architectures and is ranked at the top 1 accuracy on KITTI stereo 2012, 2015 and Middlebury benchmarks, as well as the top 1 on SceneFlow dataset with a substantial improvement on the size of the network and the speed of inference. The code is available at https://github.com/XuelianCheng/LEAStereo.

Markus Hiller, Krista A. Ehinger, Tom Drummond

We present a novel bi-directional Transformer architecture (BiXT) which scales linearly with input size in terms of computational cost and memory consumption, but does not suffer the drop in performance or limitation to only one input modality seen with other efficient Transformer-based approaches. BiXT is inspired by the Perceiver architectures but replaces iterative attention with an efficient bi-directional cross-attention module in which input tokens and latent variables attend to each other simultaneously, leveraging a naturally emerging attention-symmetry between the two. This approach unlocks a key bottleneck experienced by Perceiver-like architectures and enables the processing and interpretation of both semantics ('what') and location ('where') to develop alongside each other over multiple layers -- allowing its direct application to dense and instance-based tasks alike. By combining efficiency with the generality and performance of a full Transformer architecture, BiXT can process longer sequences like point clouds, text or images at higher feature resolutions and achieves competitive performance across a range of tasks like point cloud part segmentation, semantic image segmentation, image classification, hierarchical sequence modeling and document retrieval. Our experiments demonstrate that BiXT models outperform larger competitors by leveraging longer sequences more efficiently on vision tasks like classification and segmentation, and perform on par with full Transformer variants on sequence modeling and document retrieval -- but require $28\%$ fewer FLOPs and are up to $8.4\times$ faster.

Fan Jiang, Tom Drummond, Trevor Cohn

Cross-lingual open domain question answering (CLQA) is a complex problem, comprising cross-lingual retrieval from a multilingual knowledge base, followed by answer generation in the query language. Both steps are usually tackled by separate models, requiring substantial annotated datasets, and typically auxiliary resources, like machine translation systems to bridge between languages. In this paper, we show that CLQA can be addressed using a single encoder-decoder model. To effectively train this model, we propose a self-supervised method based on exploiting the cross-lingual link structure within Wikipedia. We demonstrate how linked Wikipedia pages can be used to synthesise supervisory signals for cross-lingual retrieval, through a form of cloze query, and generate more natural questions to supervise answer generation. Together, we show our approach, \texttt{CLASS}, outperforms comparable methods on both supervised and zero-shot language adaptation settings, including those using machine translation.

Haopeng Li, Qiuhong Ke, Mingming Gong et al.

While significant advancements have been made in video question answering (VideoQA), the potential benefits of enhancing model generalization through tailored difficulty scheduling have been largely overlooked in existing research. This paper seeks to bridge that gap by incorporating VideoQA into a curriculum learning (CL) framework that progressively trains models from simpler to more complex data. Recognizing that conventional self-paced CL methods rely on training loss for difficulty measurement, which might not accurately reflect the intricacies of video-question pairs, we introduce the concept of uncertainty-aware CL. Here, uncertainty serves as the guiding principle for dynamically adjusting the difficulty. Furthermore, we address the challenge posed by uncertainty by presenting a probabilistic modeling approach for VideoQA. Specifically, we conceptualize VideoQA as a stochastic computation graph, where the hidden representations are treated as stochastic variables. This yields two distinct types of uncertainty: one related to the inherent uncertainty in the data and another pertaining to the model's confidence. In practice, we seamlessly integrate the VideoQA model into our framework and conduct comprehensive experiments. The findings affirm that our approach not only achieves enhanced performance but also effectively quantifies uncertainty in the context of VideoQA.

Fan Jiang, Tom Drummond, Trevor Cohn

Recent approaches to multilingual open-domain question answering (MLODQA) have achieved promising results given abundant language-specific training data. However, the considerable annotation cost limits the application of these methods for underrepresented languages. We introduce a \emph{few-shot learning} approach to synthesise large-scale multilingual data from large language models (LLMs). Our method begins with large-scale self-supervised pre-training using WikiData, followed by training on high-quality synthetic multilingual data generated by prompting LLMs with few-shot supervision. The final model, \textsc{FsModQA}, significantly outperforms existing few-shot and supervised baselines in MLODQA and cross-lingual and monolingual retrieval. We further show our method can be extended for effective zero-shot adaptation to new languages through a \emph{cross-lingual prompting} strategy with only English-supervised data, making it a general and applicable solution for MLODQA tasks without costly large-scale annotation.

Haopeng Li, Tom Drummond, Mingming Gong et al.

Significant progress has been made in the field of video question answering (VideoQA) thanks to deep learning and large-scale pretraining. Despite the presence of sophisticated model structures and powerful video-text foundation models, most existing methods focus solely on maximizing the correlation between answers and video-question pairs during training. We argue that these models often establish shortcuts, resulting in spurious correlations between questions and answers, especially when the alignment between video and text data is suboptimal. To address these spurious correlations, we propose a novel training framework in which the model is compelled to acknowledge its ignorance when presented with an intervened question, rather than making guesses solely based on superficial question-answer correlations. We introduce methodologies for intervening in questions, utilizing techniques such as displacement and perturbation, and design frameworks for the model to admit its lack of knowledge in both multi-choice VideoQA and open-ended settings. In practice, we integrate a state-of-the-art model into our framework to validate its effectiveness. The results clearly demonstrate that our framework can significantly enhance the performance of VideoQA models with minimal structural modifications.

Bryce Ferenczi, Michael Burke, Tom Drummond

Various works have aimed at combining the inference efficiency of recurrent models and training parallelism of multi-head attention for sequence modeling. However, most of these works focus on tasks with fixed-dimension observation spaces, such as individual tokens in language modeling or pixels in image completion. To handle an observation space of varying size, we propose a novel algorithm that alternates between cross-attention between a 2D latent state and observation, and a discounted cumulative sum over the sequence dimension to efficiently accumulate historical information. We find this resampling cycle is critical for performance. To evaluate efficient sequence modeling in this domain, we introduce two multi-agent intention tasks: simulated agents chasing bouncing particles and micromanagement analysis in professional StarCraft II games. Our algorithm achieves comparable accuracy with a lower parameter count, faster training and inference compared to existing methods.

Li Haopeng, Ke Qiuhong, Gong Mingming et al.

Modern video summarization methods are based on deep neural networks that require a large amount of annotated data for training. However, existing datasets for video summarization are small-scale, easily leading to over-fitting of the deep models. Considering that the annotation of large-scale datasets is time-consuming, we propose a multimodal self-supervised learning framework to obtain semantic representations of videos, which benefits the video summarization task. Specifically, the self-supervised learning is conducted by exploring the semantic consistency between the videos and text in both coarse-grained and fine-grained fashions, as well as recovering masked frames in the videos. The multimodal framework is trained on a newly-collected dataset that consists of video-text pairs. Additionally, we introduce a progressive video summarization method, where the important content in a video is pinpointed progressively to generate better summaries. Extensive experiments have proved the effectiveness and superiority of our method in rank correlation coefficients and F-score.

Rongkai Ma, Pengfei Fang, Gil Avraham et al.

Learning and generalizing to novel concepts with few samples (Few-Shot Learning) is still an essential challenge to real-world applications. A principle way of achieving few-shot learning is to realize a model that can rapidly adapt to the context of a given task. Dynamic networks have been shown capable of learning content-adaptive parameters efficiently, making them suitable for few-shot learning. In this paper, we propose to learn the dynamic kernels of a convolution network as a function of the task at hand, enabling faster generalization. To this end, we obtain our dynamic kernels based on the entire task and each sample and develop a mechanism further conditioning on each individual channel and position independently. This results in dynamic kernels that simultaneously attend to the global information whilst also considering minuscule details available. We empirically show that our model improves performance on few-shot classification and detection tasks, achieving a tangible improvement over several baseline models. This includes state-of-the-art results on 4 few-shot classification benchmarks: mini-ImageNet, tiered-ImageNet, CUB and FC100 and competitive results on a few-shot detection dataset: MS COCO-PASCAL-VOC.

Rongkai Ma, Pengfei Fang, Tom Drummond et al.

Learning and generalizing from limited examples, i,e, few-shot learning, is of core importance to many real-world vision applications. A principal way of achieving few-shot learning is to realize an embedding where samples from different classes are distinctive. Recent studies suggest that embedding via hyperbolic geometry enjoys low distortion for hierarchical and structured data, making it suitable for few-shot learning. In this paper, we propose to learn a context-aware hyperbolic metric to characterize the distance between a point and a set associated with a learned set to set distance. To this end, we formulate the metric as a weighted sum on the tangent bundle of the hyperbolic space and develop a mechanism to obtain the weights adaptively and based on the constellation of the points. This not only makes the metric local but also dependent on the task in hand, meaning that the metric will adapt depending on the samples that it compares. We empirically show that such metric yields robustness in the presence of outliers and achieves a tangible improvement over baseline models. This includes the state-of-the-art results on five popular few-shot classification benchmarks, namely mini-ImageNet, tiered-ImageNet, Caltech-UCSD Birds-200-2011 (CUB), CIFAR-FS, and FC100.

Tianyu Zhu, Rongkai Ma, Mehrtash Harandi et al.

Tracking requires building a discriminative model for the target in the inference stage. An effective way to achieve this is online learning, which can comfortably outperform models that are only trained offline. Recent research shows that visual tracking benefits significantly from the unification of visual tracking and segmentation due to its pixel-level discrimination. However, it imposes a great challenge to perform online learning for such a unified model. A segmentation model cannot easily learn from prior information given in the visual tracking scenario. In this paper, we propose TrackMLP: a novel meta-learning method optimized to learn from only partial information to resolve the imposed challenge. Our model is capable of extensively exploiting limited prior information hence possesses much stronger target-background discriminability than other online learning methods. Empirically, we show that our model achieves state-of-the-art performance and tangible improvement over competing models. Our model achieves improved average overlaps of66.0%,67.1%, and68.5% in VOT2019, VOT2018, and VOT2016 datasets, which are 6.4%,7.3%, and6.4% higher than our baseline. Code will be made publicly available.

Lie Ju, Xin Wang, Lin Wang et al.

In the real world, medical datasets often exhibit a long-tailed data distribution (i.e., a few classes occupy most of the data, while most classes have rarely few samples), which results in a challenging imbalance learning scenario. For example, there are estimated more than 40 different kinds of retinal diseases with variable morbidity, however with more than 30+ conditions are very rare from the global patient cohorts, which results in a typical long-tailed learning problem for deep learning-based screening models. In this study, we propose class subset learning by dividing the long-tailed data into multiple class subsets according to prior knowledge, such as regions and phenotype information. It enforces the model to focus on learning the subset-specific knowledge. More specifically, there are some relational classes that reside in the fixed retinal regions, or some common pathological features are observed in both the majority and minority conditions. With those subsets learnt teacher models, then we are able to distill the multiple teacher models into a unified model with weighted knowledge distillation loss. The proposed framework proved to be effective for the long-tailed retinal diseases recognition task. The experimental results on two different datasets demonstrate that our method is flexible and can be easily plugged into many other state-of-the-art techniques with significant improvements.

Morris Gu, Akansel Cosgun, Wesley P. Chan et al.

We present an approach for visualizing mobile robots through an Augmented Reality headset when there is no line-of-sight visibility between the robot and the human. Three elements are visualized in Augmented Reality: 1) Robot's 3D model to indicate its position, 2) An arrow emanating from the robot to indicate its planned movement direction, and 3) A 2D grid to represent the ground plane. We conduct a user study with 18 participants, in which each participant are asked to retrieve objects, one at a time, from stations at the two sides of a T-junction at the end of a hallway where a mobile robot is roaming. The results show that visualizations improved the perceived safety and efficiency of the task and led to participants being more comfortable with the robot within their personal spaces. Furthermore, visualizing the motion intent in addition to the robot model was found to be more effective than visualizing the robot model alone. The proposed system can improve the safety of automated warehouses by increasing the visibility and predictability of robots.

Benjamin Waymouth, Akansel Cosgun, Rhys Newbury et al.

An appropriate user interface to collect human demonstration data for deformable object manipulation has been mostly overlooked in the literature. We present an interaction design for demonstrating cloth folding to robots. Users choose pick and place points on the cloth and can preview a visualization of a simulated cloth before real-robot execution. Two interfaces are proposed: A 2D display-and-mouse interface where points are placed by clicking on an image of the cloth, and a 3D Augmented Reality interface where the chosen points are placed by hand gestures. We conduct a user study with 18 participants, in which each user completed two sequential folds to achieve a cloth goal shape. Results show that while both interfaces were acceptable, the 3D interface was found to be more suitable for understanding the task, and the 2D interface suitable for repetition. Results also found that fold previews improve three key metrics: task efficiency, the ability to predict the final shape of the cloth and overall user satisfaction.

Tianyu Zhu, Markus Hiller, Mahsa Ehsanpour et al.

Tracking a time-varying indefinite number of objects in a video sequence over time remains a challenge despite recent advances in the field. Most existing approaches are not able to properly handle multi-object tracking challenges such as occlusion, in part because they ignore long-term temporal information. To address these shortcomings, we present MO3TR: a truly end-to-end Transformer-based online multi-object tracking (MOT) framework that learns to handle occlusions, track initiation and termination without the need for an explicit data association module or any heuristics. MO3TR encodes object interactions into long-term temporal embeddings using a combination of spatial and temporal Transformers, and recursively uses the information jointly with the input data to estimate the states of all tracked objects over time. The spatial attention mechanism enables our framework to learn implicit representations between all the objects and the objects to the measurements, while the temporal attention mechanism focuses on specific parts of past information, allowing our approach to resolve occlusions over multiple frames. Our experiments demonstrate the potential of this new approach, achieving results on par with or better than the current state-of-the-art on multiple MOT metrics for several popular multi-object tracking benchmarks.

Rhys Newbury, Akansel Cosgun, Tysha Crowley-Davis et al.

Humans are highly skilled in communicating their intent for when and where a handover would occur. However, even the state-of-the-art robotic implementations for handovers typically lack of such communication skills. This study investigates visualization of the robot's internal state and intent for Human-to-Robot Handovers using Augmented Reality. Specifically, we explore the use of visualized 3D models of the object and the robotic gripper to communicate the robot's estimation of where the object is and the pose in which the robot intends to grasp the object. We tested this design via a user study with 16 participants, in which each participant handed over a cube-shaped object to the robot 12 times. Results show communicating robot intent via augmented reality substantially improves the perceived experience of the users for handovers. Results also indicate that the effectiveness of augmented reality is even more pronounced for the perceived safety and fluency of the interaction when the robot makes errors in localizing the object.

Lie Ju, Xin Wang, Lin Wang et al.

Deep neural networks are known to be data-driven and label noise can have a marked impact on model performance. Recent studies have shown great robustness to classic image recognition even under a high noisy rate. In medical applications, learning from datasets with label noise is more challenging since medical imaging datasets tend to have asymmetric (class-dependent) noise and suffer from high observer variability. In this paper, we systematically discuss and define the two common types of label noise in medical images - disagreement label noise from inconsistency expert opinions and single-target label noise from wrong diagnosis record. We then propose an uncertainty estimation-based framework to handle these two label noise amid the medical image classification task. We design a dual-uncertainty estimation approach to measure the disagreement label noise and single-target label noise via Direct Uncertainty Prediction and Monte-Carlo-Dropout. A boosting-based curriculum training procedure is later introduced for robust learning. We demonstrate the effectiveness of our method by conducting extensive experiments on three different diseases: skin lesions, prostate cancer, and retinal diseases. We also release a large re-engineered database that consists of annotations from more than ten ophthalmologists with an unbiased golden standard dataset for evaluation and benchmarking.

Lie Ju, Xin Wang, Xin Zhao et al.

Recently, ultra-widefield (UWF) 200\degree~fundus imaging by Optos cameras has gradually been introduced because of its broader insights for detecting more information on the fundus than regular 30 degree - 60 degree fundus cameras. Compared with UWF fundus images, regular fundus images contain a large amount of high-quality and well-annotated data. Due to the domain gap, models trained by regular fundus images to recognize UWF fundus images perform poorly. Hence, given that annotating medical data is labor intensive and time consuming, in this paper, we explore how to leverage regular fundus images to improve the limited UWF fundus data and annotations for more efficient training. We propose the use of a modified cycle generative adversarial network (CycleGAN) model to bridge the gap between regular and UWF fundus and generate additional UWF fundus images for training. A consistency regularization term is proposed in the loss of the GAN to improve and regulate the quality of the generated data. Our method does not require that images from the two domains be paired or even that the semantic labels be the same, which provides great convenience for data collection. Furthermore, we show that our method is robust to noise and errors introduced by the generated unlabeled data with the pseudo-labeling technique. We evaluated the effectiveness of our methods on several common fundus diseases and tasks, such as diabetic retinopathy (DR) classification, lesion detection and tessellated fundus segmentation. The experimental results demonstrate that our proposed method simultaneously achieves superior generalizability of the learned representations and performance improvements in multiple tasks.

Gil Avraham, Yan Zuo, Tom Drummond

Domain adaptation and generative modelling have collectively mitigated the expensive nature of data collection and labelling by leveraging the rich abundance of accurate, labelled data in simulation environments. In this work, we study the performance gap that exists between representations optimised for localisation on simulation environments and the application of such representations in a real-world setting. Our method exploits the shared geometric similarities between simulation and real-world environments whilst maintaining invariance towards visual discrepancies. This is achieved by optimising a representation extractor to project both simulated and real representations into a shared representation space. Our method uses a symmetrical adversarial approach which encourages the representation extractor to conceal the domain that features are extracted from and simultaneously preserves robust attributes between source and target domains that are beneficial for localisation. We evaluate our method by adapting representations optimised for indoor Habitat simulated environments (Matterport3D and Replica) to a real-world indoor environment (Active Vision Dataset), showing that it compares favourably against fully-supervised approaches.

Yan Zuo, Tom Drummond

This paper presents a novel ensemble learning approach called Residual Likelihood Forests (RLF). Our weak learners produce conditional likelihoods that are sequentially optimized using global loss in the context of previous learners within a boosting-like framework (rather than probability distributions that are measured from observed data) and are combined multiplicatively (rather than additively). This increases the efficiency of our strong classifier, allowing for the design of classifiers which are more compact in terms of model capacity. We apply our method to several machine learning classification tasks, showing significant improvements in performance. When compared against several ensemble approaches including Random Forests and Gradient Boosted Trees, RLFs offer a significant improvement in performance whilst concurrently reducing the required model size.

Abdelhak Loukkal, Yves Grandvalet, Tom Drummond et al.

Camera-based end-to-end driving neural networks bring the promise of a low-cost system that maps camera images to driving control commands. These networks are appealing because they replace laborious hand engineered building blocks but their black-box nature makes them difficult to delve in case of failure. Recent works have shown the importance of using an explicit intermediate representation that has the benefits of increasing both the interpretability and the accuracy of networks' decisions. Nonetheless, these camera-based networks reason in camera view where scale is not homogeneous and hence not directly suitable for motion forecasting. In this paper, we introduce a novel monocular camera-only holistic end-to-end trajectory planning network with a Bird-Eye-View (BEV) intermediate representation that comes in the form of binary Occupancy Grid Maps (OGMs). To ease the prediction of OGMs in BEV from camera images, we introduce a novel scheme where the OGMs are first predicted as semantic masks in camera view and then warped in BEV using the homography between the two planes. The key element allowing this transformation to be applied to 3D objects such as vehicles, consists in predicting solely their footprint in camera-view, hence respecting the flat world hypothesis implied by the homography.

Shray Bansal, Rhys Newbury, Wesley Chan et al.

The increasing presence of robots alongside humans, such as in human-robot teams in manufacturing, gives rise to research questions about the kind of behaviors people prefer in their robot counterparts. We term actions that support interaction by reducing future interference with others as supportive robot actions and investigate their utility in a co-located manipulation scenario. We compare two robot modes in a shared table pick-and-place task: (1) Task-oriented: the robot only takes actions to further its own task objective and (2) Supportive: the robot sometimes prefers supportive actions to task-oriented ones when they reduce future goal-conflicts. Our experiments in simulation, using a simplified human model, reveal that supportive actions reduce the interference between agents, especially in more difficult tasks, but also cause the robot to take longer to complete the task. We implemented these modes on a physical robot in a user study where a human and a robot perform object placement on a shared table. Our results show that a supportive robot was perceived as a more favorable coworker by the human and also reduced interference with the human in the more difficult of two scenarios. However, it also took longer to complete the task highlighting an interesting trade-off between task-efficiency and human-preference that needs to be considered before designing robot behavior for close-proximity manipulation scenarios.

Rhys Newbury, Kerry He, Akansel Cosgun et al.

We study the problem of placing a grasped object on an empty flat surface in an upright orientation, such as placing a cup on its bottom rather than on its side. We aim to find the required object rotation such that when the gripper is opened after the object makes contact with the surface, the object would be stably placed in the upright orientation. We iteratively use two neural networks. At every iteration, we use a convolutional neural network to estimate the required object rotation, which is executed by the robot, and then a separate convolutional neural network to estimate the quality of a placement in its current orientation. Our approach places previously unseen objects in upright orientations with a success rate of 98.1% in free space and 90.3% with a simulated robotic arm, using a dataset of 50 everyday objects in simulation experiments. Real-world experiments were performed, which achieved an 88.0% success rate, which serves as a proof-of-concept for direct sim-to-real transfer.

Lie Ju, Xin Wang, Quan Zhou et al.

For decades, advances in retinal imaging technology have enabled effective diagnosis and management of retinal disease using fundus cameras. Recently, ultra-wide-field (UWF) fundus imaging by Optos camera is gradually put into use because of its broader insights on fundus for some lesions that are not typically seen in traditional fundus images. Research on traditional fundus images is an active topic but studies on UWF fundus images are few. One of the most important reasons is that UWF fundus images are hard to obtain. In this paper, for the first time, we explore domain adaptation from the traditional fundus to UWF fundus images. We propose a flexible framework to bridge the domain gap between two domains and co-train a UWF fundus diagnosis model by pseudo-labelling and adversarial learning. We design a regularisation technique to regulate the domain adaptation. Also, we apply MixUp to overcome the over-fitting issue from incorrect generated pseudo-labels. Our experimental results on either single or both domains demonstrate that the proposed method can well adapt and transfer the knowledge from traditional fundus images to UWF fundus images and improve the performance of retinal disease recognition.

Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza et al.

Event cameras are paradigm-shifting novel sensors that report asynchronous, per-pixel brightness changes called 'events' with unparalleled low latency. This makes them ideal for high speed, high dynamic range scenes where conventional cameras would fail. Recent work has demonstrated impressive results using Convolutional Neural Networks (CNNs) for video reconstruction and optic flow with events. We present strategies for improving training data for event based CNNs that result in 20-40% boost in performance of existing state-of-the-art (SOTA) video reconstruction networks retrained with our method, and up to 15% for optic flow networks. A challenge in evaluating event based video reconstruction is lack of quality ground truth images in existing datasets. To address this, we present a new High Quality Frames (HQF) dataset, containing events and ground truth frames from a DAVIS240C that are well-exposed and minimally motion-blurred. We evaluate our method on HQF + several existing major event camera datasets.

Luke Ditria, Benjamin J. Meyer, Tom Drummond

Many existing conditional Generative Adversarial Networks (cGANs) are limited to conditioning on pre-defined and fixed class-level semantic labels or attributes. We propose an open set GAN architecture (OpenGAN) that is conditioned per-input sample with a feature embedding drawn from a metric space. Using a state-of-the-art metric learning model that encodes both class-level and fine-grained semantic information, we are able to generate samples that are semantically similar to a given source image. The semantic information extracted by the metric learning model transfers to out-of-distribution novel classes, allowing the generative model to produce samples that are outside of the training distribution. We show that our proposed method is able to generate 256$\times$256 resolution images from novel classes that are of similar visual quality to those from the training classes. In lieu of a source image, we demonstrate that random sampling of the metric space also results in high-quality samples. We show that interpolation in the feature space and latent space results in semantically and visually plausible transformations in the image space. Finally, the usefulness of the generated samples to the downstream task of data augmentation is demonstrated. We show that classifier performance can be significantly improved by augmenting the training data with OpenGAN samples on classes that are outside of the GAN training distribution.

Luis Guerra, Bohan Zhuang, Ian Reid et al.

Instantaneous and on demand accuracy-efficiency trade-off has been recently explored in the context of neural networks slimming. In this paper, we propose a flexible quantization strategy, termed Switchable Precision neural Networks (SP-Nets), to train a shared network capable of operating at multiple quantization levels. At runtime, the network can adjust its precision on the fly according to instant memory, latency, power consumption and accuracy demands. For example, by constraining the network weights to 1-bit with switchable precision activations, our shared network spans from BinaryConnect to Binarized Neural Network, allowing to perform dot-products using only summations or bit operations. In addition, a self-distillation scheme is proposed to increase the performance of the quantized switches. We tested our approach with three different quantizers and demonstrate the performance of SP-Nets against independently trained quantized models in classification accuracy for Tiny ImageNet and ImageNet datasets using ResNet-18 and MobileNet architectures.

Luis Guerra, Bohan Zhuang, Ian Reid et al.

Neural network quantization and pruning are two techniques commonly used to reduce the computational complexity and memory footprint of these models for deployment. However, most existing pruning strategies operate on full-precision and cannot be directly applied to discrete parameter distributions after quantization. In contrast, we study a combination of these two techniques to achieve further network compression. In particular, we propose an effective pruning strategy for selecting redundant low-precision filters. Furthermore, we leverage Bayesian optimization to efficiently determine the pruning ratio for each layer. We conduct extensive experiments on CIFAR-10 and ImageNet with various architectures and precisions. In particular, for ResNet-18 on ImageNet, we prune 26.12% of the model size with Binarized Neural Network quantization, achieving a top-1 classification accuracy of 47.32% in a model of 2.47 MB and 59.30% with a 2-bit DoReFa-Net in 4.36 MB.

Yunyan Xing, Zongyuan Ge, Rui Zeng et al.

Recent works show that Generative Adversarial Networks (GANs) can be successfully applied to chest X-ray data augmentation for lung disease recognition. However, the implausible and distorted pathology features generated from the less than perfect generator may lead to wrong clinical decisions. Why not keep the original pathology region? We proposed a novel approach that allows our generative model to generate high quality plausible images that contain undistorted pathology areas. The main idea is to design a training scheme based on an image-to-image translation network to introduce variations of new lung features around the pathology ground-truth area. Moreover, our model is able to leverage both annotated disease images and unannotated healthy lung images for the purpose of generation. We demonstrate the effectiveness of our model on two tasks: (i) we invite certified radiologists to assess the quality of the generated synthetic images against real and other state-of-the-art generative models, and (ii) data augmentation to improve the performance of disease localisation.

Gil Avraham, Yan Zuo, Thanuja Dharmasiri et al.

Continuously estimating an agent's state space and a representation of its surroundings has proven vital towards full autonomy. A shared common ground among systems which successfully achieve this feat is the integration of previously encountered observations into the current state being estimated. This necessitates the use of a memory module for incorporating previously visited states whilst simultaneously offering an internal representation of the observed environment. In this work we develop a memory module which contains rigidly aligned point-embeddings that represent a coherent scene structure acquired from an RGB-D sequence of observations. The point-embeddings are extracted using modern convolutional neural network architectures, and alignment is performed by computing a dense correspondence matrix between a new observation and the current embeddings residing in the memory module. The whole framework is end-to-end trainable, resulting in a recurrent joint optimisation of the point-embeddings contained in the memory. This process amplifies the shared information across states, providing increased robustness and accuracy. We show significant improvement of our method across a set of experiments performed on the synthetic VIZDoom environment and a real world Active Vision Dataset.

Akansel Cosgun, Luke Ditria, Shayne D'Lima et al.

Grasping is the dominant approach for robot manipulation, but only a single object can be grasped at a time. Nonprehensile manipulation offers richer set of interactions, however state-of-the-art is limited to using the end-effector only. We propose using a robot link (forearm) to push multiple objects at once. In a simulated task where the robot's task is to sort two kinds of objects into their respective goal regions, we show that a greedy strategy that uses a combination of forearm pushes and pick and place operations reduces task completion time by %28 compared to picking and placing each object individually.

Akansel Cosgun, Thomas Rowntree, Ian Reid et al.

Robots need to learn behaviors in intuitive and practical ways for widespread deployment in human environments. To learn a robot behavior end-to-end, we train a variant of the ResNet that maps eye-in-hand camera images to end-effector velocities. In our setup, a human teacher demonstrates the task via joystick. We show that a simple servoing task can be learned in less than an hour including data collection, model training and deployment time. Moreover, 16 minutes of demonstrations were enough for the robot to learn the task.

Rhys Newbury, Akansel Cosgun, Mehmet Koseoglu et al.

We present a robotic system capable of navigating autonomously by following a line and taking good quality pictures of people. When a group of people are detected, the robot rotates towards them and then back to line while continuously taking pictures from different angles. Each picture is processed in the cloud where its quality is estimated in a two-stage algorithm. First, features such as the face orientation and likelihood of facial emotions are input to a fully connected neural network to assign a quality score to each face. Second, a representation is extracted by abstracting faces from the image and it is input to a to Convolutional Neural Network (CNN) to classify the quality of the overall picture. We collected a dataset in which a picture was labeled as good quality if subjects are well-positioned in the image and oriented towards the camera with a pleasant expression. Our approach detected the quality of pictures with 78.4% accuracy in this dataset and received a better mean user rating (3.71/5) than a heuristic method that uses photographic composition procedures in a study where 97 human judges rated each picture. A statistical analysis against the state-of-the-art verified the quality of the resulting pictures.

Timo Stoffregen, Guillermo Gallego, Tom Drummond et al.

In contrast to traditional cameras, whose pixels have a common exposure time, event-based cameras are novel bio-inspired sensors whose pixels work independently and asynchronously output intensity changes (called "events"), with microsecond resolution. Since events are caused by the apparent motion of objects, event-based cameras sample visual information based on the scene dynamics and are, therefore, a more natural fit than traditional cameras to acquire motion, especially at high speeds, where traditional cameras suffer from motion blur. However, distinguishing between events caused by different moving objects and by the camera's ego-motion is a challenging task. We present the first per-event segmentation method for splitting a scene into independently moving objects. Our method jointly estimates the event-object associations (i.e., segmentation) and the motion parameters of the objects (or the background) by maximization of an objective function, which builds upon recent results on event-based motion-compensation. We provide a thorough evaluation of our method on a public dataset, outperforming the state-of-the-art by as much as 10%. We also show the first quantitative evaluation of a segmentation algorithm for event cameras, yielding around 90% accuracy at 4 pixels relative displacement.

Benjamin J. Meyer, Tom Drummond

State-of-the-art deep neural network recognition systems are designed for a static and closed world. It is usually assumed that the distribution at test time will be the same as the distribution during training. As a result, classifiers are forced to categorise observations into one out of a set of predefined semantic classes. Robotic problems are dynamic and open world; a robot will likely observe objects that are from outside of the training set distribution. Classifier outputs in robotic applications can lead to real-world robotic action and as such, a practical recognition system should not silently fail by confidently misclassifying novel observations. We show how a deep metric learning classification system can be applied to such open set recognition problems, allowing the classifier to label novel observations as unknown. Further to detecting novel examples, we propose an open set active learning approach that allows a robot to efficiently query a user about unknown observations. Our approach enables a robot to improve its understanding of the true distribution of data in the environment, from a small number of label queries. Experimental results show that our approach significantly outperforms comparable methods in both the open set recognition and active learning problems.

Sourav Garg, Madhu Babu, Thanuja Dharmasiri et al.

Visual place recognition (VPR) - the act of recognizing a familiar visual place - becomes difficult when there is extreme environmental appearance change or viewpoint change. Particularly challenging is the scenario where both phenomena occur simultaneously, such as when returning for the first time along a road at night that was previously traversed during the day in the opposite direction. While such problems can be solved with panoramic sensors, humans solve this problem regularly with limited field of view vision and without needing to constantly turn around. In this paper, we present a new depth- and temporal-aware visual place recognition system that solves the opposing viewpoint, extreme appearance-change visual place recognition problem. Our system performs sequence-to-single matching by extracting depth-filtered keypoints using a state-of-the-art depth estimation pipeline, constructing a keypoint sequence over multiple frames from the reference dataset, and comparing those keypoints to those in a single query image. We evaluate the system on a challenging benchmark dataset and show that it consistently outperforms state-of-the-art techniques. We also develop a range of diagnostic simulation experiments that characterize the contribution of depth-filtered keypoint sequences with respect to key domain parameters including degree of appearance change and camera motion.

Yan Zuo, Gil Avraham, Tom Drummond

The practice of transforming raw data to a feature space so that inference can be performed in that space has been popular for many years. Recently, rapid progress in deep neural networks has given both researchers and practitioners enhanced methods that increase the richness of feature representations, be it from images, text or speech. In this work we show how a constructed latent space can be explored in a controlled manner and argue that this complements well founded inference methods. For constructing the latent space a Variational Autoencoder is used. We present a novel controller module that allows for smooth traversal in the latent space and construct an end-to-end trainable framework. We explore the applicability of our method for performing spatial transformations as well as kinematics for predicting future latent vectors of a video sequence.