A5000069758- Advanced Neural Network Applications
- Robotics and Sensor-Based Localization
- Robotic Path Planning Algorithms
- Anomaly Detection Techniques and Applications
- Video Surveillance and Tracking Methods
- Advanced Vision and Imaging
- Remote Sensing and LiDAR Applications
- Advanced Image and Video Retrieval Techniques
- Multimodal Machine Learning Applications
- Statistical Methods and Bayesian Inference
- Face recognition and analysis
- Medical Imaging Techniques and Applications
- Autonomous Vehicle Technology and Safety
- Markov Chains and Monte Carlo Methods
- Medical Image Segmentation Techniques
- Human Pose and Action Recognition
- Adversarial Robustness in Machine Learning
- Industrial Vision Systems and Defect Detection
- Bayesian Methods and Mixture Models
- Radiomics and Machine Learning in Medical Imaging
- Explainable Artificial Intelligence (XAI)
- Computer Graphics and Visualization Techniques
- Image Enhancement Techniques
- Domain Adaptation and Few-Shot Learning
- Brain Tumor Detection and Classification
Massachusetts Institute of Technology
2023-2024
Carnegie Mellon University
2020-2021
Microsoft Research (United Kingdom)
2016-2018
Microsoft (United States)
2016
Traversing terrain with good traction is crucial for achieving fast off-road navigation. Instead of manually designing costs based on features, existing methods learn properties directly from data via self-supervision to automatically penalize trajectories moving through undesirable terrain, but challenges remain in properly quantifying and mitigating the risk due uncertainty learned models. To this end, we present evidential autonomy (EVORA), a unified framework uncertainty-aware model plan...
We propose the autofocus convolutional layer for semantic segmentation with objective of enhancing capabilities neural networks multi-scale processing. Autofocus layers adaptively change size effective receptive field based on processed context to generate more powerful features. This is achieved by parallelising multiple different dilation rates, combined an attention mechanism that learns focus optimal scales driven context. By sharing weights parallel convolutions we make network...
Active sensing through the use of Adaptive Depth Sensors is a nascent field, with potential in areas such as Advanced driver-assistance systems (ADAS). They do however require dynamically driving laser / light-source to specific location capture information, one class sensor being Triangulation Light Curtains (LC). In this work, we introduce novel approach that exploits prior depth distributions from RGB cameras drive Curtain's line regions uncertainty get new measurements. These...
3D object trackers usually require training on large amounts of annotated data that is expensive and time-consuming to collect. Instead, we propose leveraging vast unlabeled datasets by self-supervised metric learning trackers, with a focus association. Large scale annotations for are cheaply obtained automatic detection association across frames. We show how these can be used in principled manner learn point-cloud embeddings effective tracking. estimate incorporate uncertainty tracking more...
To safely navigate unknown environments, robots must accurately perceive dynamic obstacles.Instead of directly measuring the scene depth with a LiDAR sensor, we explore use much cheaper and higher resolution sensor: programmable light curtains.Light curtains are controllable sensors that sense only along surface user selects.We to estimate safety envelope scene: hypothetical separates robot from all obstacles.We show generating random locations (from particular distribution) can quickly...
Markov chain Monte Carlo (MCMC) is one of the main workhorses probabilistic inference, but it notoriously hard to measure quality approximate posterior samples. This challenge particularly salient in black box inference methods, which can hide details and obscure failures. In this work, we extend recently introduced bidirectional technique evaluate MCMC-based algorithms. By running annealed importance sampling (AIS) chains both from prior vice versa on simulated data, upper bound expectation...
As factories continue to evolve into collaborative spaces with multiple robots working together human supervisors in the loop, ensuring safety for all actors involved becomes critical. Currently, laser-based light curtain sensors are widely used monitoring. While these conventional meet high accuracy standards, they difficult reconfigure and can only monitor a fixed user-defined region of space. Furthermore, typically expensive. Instead, we leverage controllable depth sensor, programmable...
Current state-of-the-art trackers often fail due to distractorsand large object appearance changes. In this work, we explore the use ofdense optical flow improve tracking robustness. Our main insight is that, because estimation can also have errors, need incorporate an estimate of uncertainty for robust tracking. We present a novel framework which combines and information track objects in challenging scenarios. experimentally verify that our improves robustness, leading new results. Further,...
To navigate in an environment safely and autonomously, robots must accurately estimate where obstacles are how they move.Instead of using expensive traditional 3D sensors, we explore the use a much cheaper, faster, higher resolution alternative: programmable light curtains.Light curtains controllable depth sensor that sense only along surface user selects.We adapt probabilistic method based on particle filters occupancy grids to explicitly position velocity points scene partial measurements...
Deep learning object detectors often return false positives with very high confidence. Although they optimize generic detection performance, such as mean average precision (mAP), are not designed for reliability. For a reliable system, if confidence is made, we would want certainty that the has indeed been detected. To achieve this, have developed set of verification tests which proposed must pass to be accepted. We develop theoretical framework proves that, under certain assumptions, our...
Traversing terrain with good traction is crucial for achieving fast off-road navigation. Instead of manually designing costs based on features, existing methods learn properties directly from data via self-supervision to automatically penalize trajectories moving through undesirable terrain, but challenges remain properly quantify and mitigate the risk due uncertainty in learned models. To this end, work proposes a unified framework uncertainty-aware model plan risk-aware trajectories. For...
To navigate in an environment safely and autonomously, robots must accurately estimate where obstacles are how they move. Instead of using expensive traditional 3D sensors, we explore the use a much cheaper, faster, higher resolution alternative: programmable light curtains. Light curtains controllable depth sensor that sense only along surface user selects. We adapt probabilistic method based on particle filters occupancy grids to explicitly position velocity points scene partial...
3D object trackers usually require training on large amounts of annotated data that is expensive and time-consuming to collect. Instead, we propose leveraging vast unlabeled datasets by self-supervised metric learning trackers, with a focus association. Large scale annotations for are cheaply obtained automatic detection association across frames. We show how these can be used in principled manner learn point-cloud embeddings effective tracking. estimate incorporate uncertainty tracking more...