Dropout Variational Inference, or Sampling, has been recently proposed as an approximation technique for Bayesian Deep Learning and evaluated image classification regression tasks. This paper investigates the utility of Sampling object detection first time. We demonstrate how label uncertainty can be extracted from a state-of-the-art system via Sampling. evaluate this approach on large synthetic dataset 30,000 images, real-world captured by mobile robot in versatile campus environment. show...

We introduce Probabilistic Object Detection, the task of detecting objects in images and accurately quantifying spatial semantic uncertainties detections. Given lack methods capable assessing such probabilistic object detections, we present new Probability-based Detection Quality measure (PDQ). Unlike AP-based measures, PDQ has no arbitrary thresholds rewards label quality, foreground/background separation quality while explicitly penalising false positive negative contrast with existing mAP...

There has been a recent emergence of sampling-based techniques for estimating epistemic uncertainty in deep neural networks. While these methods can be applied to classification or semantic segmentation tasks by simply averaging samples, this is not the case object detection, where detection sample bounding boxes must accurately associated and merged. A weak merging strategy significantly degrade performance detector yield an unreliable measure. This paper provides first in-depth...

In open set recognition, deep neural networks encounter object classes that were unknown during training. Existing classifiers distinguish between known and by measuring distance in a network's logit space, assuming cluster closer to the training data than classes. However, this approach is applied post-hoc trained with cross-entropy loss, which does not guarantee clustering behaviour. To overcome limitation, we introduce Class Anchor Clustering (CAC) loss. CAC distance-based loss explicitly...

We show that ensembling effectively quantifies model uncertainty in Neural Radiance Fields (NeRFs) if a density-aware epistemic term is considered. The naive ensembles investigated prior work simply average rendered RGB images to quantify the caused by conflicting explanations of observed scene. In contrast, we additionally consider termination probabilities along individual rays identify due lack knowledge about parts scene unobserved during training. achieve new state-of-the-art...

Deployed into an open world, object detectors are prone to open-set errors, false positive detections of classes not present in the training dataset. We propose GMM-Det, a real-time method for extracting epistemic uncertainty from identify and reject errors. GMM-Det trains detector produce structured logit space that is modelled with class-specific Gaussian Mixture Models. At test time, errors identified by their low log-probability under all two common architectures, Faster R-CNN RetinaNet,...

By 2050, electric vehicles (EVs) are projected to account for 70% of global vehicle sales. While EVs provide environmental benefits, they also pose challenges energy generation, grid infrastructure, and data privacy. Current research on EV routing charge management often overlooks privacy when predicting demands, leaving sensitive mobility vulnerable. To address this, we developed a Federated Learning Transformer Network (FLTN) predict EVs' next location with enhanced measures. Each operates...

We address the problem of out-of-distribution (OOD) detection for task object detection. show that residual convolutional layers with batch normalisation produce Sensitivity-Aware FEatures (SAFE) are consistently powerful distinguishing in-distribution from detections. extract SAFE vectors every detected object, and train a multilayer perceptron on surrogate adversarially perturbed clean examples. This circumvents need realistic OOD training data, computationally expensive generative models,...

In object detection, false negatives arise when a detector fails to detect target object. To understand why detectors produce negatives, we identify five 'false negative mechanisms', where each mechanism describes how specific component inside the architecture failed. Focusing on two-stage and one-stage anchor-box architectures, introduce framework for quantifying these mechanisms. Using this framework, investigate Faster R-CNN RetinaNet fail objects in benchmark vision datasets robotics...

A false negative in object detection describes an that was not correctly localised and classified by a detector. In prior work, we introduced five 'false mechanisms' identify the specific component inside detector architecture failed to detect object. Using these mechanisms, explore how different computer vision datasets their inherent characteristics can influence failures. Specifically, investigate mechanisms of Faster R-CNN RetinaNet across datasets, namely Microsoft COCO, Pascal VOC,...

State-of-the-art lidar place recognition models exhibit unreliable performance when tested on environments different from their training dataset, which limits use in complex and evolving environments. To address this issue, we investigate the task of uncertainty-aware recognition, where each predicted must have an associated uncertainty that can be used to identify reject incorrect predictions. We introduce a novel evaluation protocol present first comprehensive benchmark for task, testing...

There are strong incentives to build models that demonstrate outstanding predictive performance on various datasets and benchmarks. We believe these risk a narrow focus the metrics used evaluate compare them -- resulting in growing body of literature metrics. This paper strives for more balanced perspective classifier by highlighting their distributions under different uncertainty showing how this can easily eclipse differences empirical classifiers. begin emphasising fundamentally discrete...

We address the challenging problem of open world object detection (OWOD), where detectors must identify objects from known classes while also identifying and continually learning to detect novel objects. Prior work has resulted in that have a relatively low ability objects, high likelihood classifying as one classes. approach by three main challenges OWOD presents introduce OW-RCNN, an detector addresses each these challenges. OW-RCNN establishes new state art using open-world evaluation...

We introduce Probabilistic Object Detection, the task of detecting objects in images and accurately quantifying spatial semantic uncertainties detections. Given lack methods capable assessing such probabilistic object detections, we present new Probability-based Detection Quality measure (PDQ).Unlike AP-based measures, PDQ has no arbitrary thresholds rewards label quality, foreground/background separation quality while explicitly penalising false positive negative contrast with existing mAP...

Non-verbal communication is essential for the social inclusion of individuals with an intellectual disability, affecting interactions others as well technological systems. This study focuses on non-symbolic people disability through generic images without specific or detailed subject matter. A key challenge in this medium discerning underlying intentions behind selected visual prompts conversation.

Are vision-language models (VLMs) open-set because they are trained on internet-scale datasets? We answer this question with a clear no - VLMs introduce closed-set assumptions via their finite query set, making them vulnerable to conditions. systematically evaluate for recognition and find frequently misclassify objects not contained in leading alarmingly low precision when tuned high recall vice versa. show that naively increasing the size of set contain more classes does mitigate problem,...

In the era of increasing concerns over cybersecurity threats, defending against backdoor attacks is paramount in ensuring integrity and reliability machine learning models. However, many existing approaches require substantial amounts data for effective mitigation, posing significant challenges practical deployment. To address this, we propose a novel approach to counter by treating their mitigation as an unlearning task. We tackle this challenge through targeted model pruning strategy,...

By 2050, global sales of electric vehicles (EVs) are predicted to account for approximately 70% all vehicle sales. However, whilst transitioning from combustion engine EVs would result in reduced carbon dioxide emissions, it place significant strain on energy generation, and grid infrastructure. Many EV studies investigated routing or charge station management, while research predicting demand at a specific location was lacking. To address this, our study focused EV's next location. We...