A5074566324- Video Surveillance and Tracking Methods
- Anomaly Detection Techniques and Applications
- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Human Pose and Action Recognition
- Modular Robots and Swarm Intelligence
- Machine Learning and Data Classification
- Robotics and Sensor-Based Localization
- Target Tracking and Data Fusion in Sensor Networks
- Impact of Light on Environment and Health
- Infrared Target Detection Methodologies
- Robot Manipulation and Learning
- Remote-Sensing Image Classification
- Robotic Path Planning Algorithms
- Visual Attention and Saliency Detection
- Advanced Optical Sensing Technologies
- Advanced Chemical Sensor Technologies
- Soft Robotics and Applications
- Face and Expression Recognition
- Fire Detection and Safety Systems
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Advanced Vision and Imaging
- Machine Learning and Algorithms
- Industrial Vision Systems and Defect Detection
Amazon (United States)
2021
Amazon (Germany)
2017-2020
The University of Adelaide
2015-2017
University of Bonn
2017
Technical University of Darmstadt
2013
Recently, very deep convolutional neural networks (CNNs) have shown outstanding performance in object recognition and also been the first choice for dense classification problems such as semantic segmentation. However, repeated subsampling operations like pooling or convolution striding CNNs lead to a significant decrease initial image resolution. Here, we present RefineNet, generic multi-path refinement network that explicitly exploits all information available along down-sampling process...
Standardized benchmarks are crucial for the majority of computer vision applications. Although leaderboards and ranking tables should not be over-claimed, often provide most objective measure performance therefore important guides reseach. Recently, a new benchmark Multiple Object Tracking, MOTChallenge, was launched with goal collecting existing data creating framework standardized evaluation multiple object tracking methods. The first release focuses on people tracking, since pedestrians...
Many recent advances in multiple target tracking aim at finding a (nearly) optimal set of trajectories within temporal window. To handle the large space possible trajectory hypotheses, it is typically reduced to finite by some form data-driven or regular discretization. In this work, we propose an alternative formulation multitarget as minimization continuous energy. Contrary approaches, focus on designing energy that corresponds more complete representation problem, rather than one amenable...
We present a novel approach to online multi-target tracking based on recurrent neural networks (RNNs). Tracking multiple objects in real-world scenes involves many challenges, including a) an a-priori unknown and time-varying number of targets, b) continuous state estimation all c) discrete combinatorial problem data association. Most previous methods involve complex models that require tedious tuning parameters. Here, we propose for the first time, end-to-end learning tracking. Existing...
In the recent past, computer vision community has developed centralized benchmarks for performance evaluation of a variety tasks, including generic object and pedestrian detection, 3D reconstruction, optical flow, single-object short-term tracking, stereo estimation. Despite potential pitfalls such benchmarks, they have proved to be extremely helpful advance state art in respective area. Interestingly, there been rather limited work on standardization quantitative multiple target tracking....
Standardized benchmarks are crucial for the majority of computer vision applications. Although leaderboards and ranking tables should not be over-claimed, often provide most objective measure performance therefore important guides research. The benchmark Multiple Object Tracking, MOTChallenge, was launched with goal to establish a standardized evaluation multiple object tracking methods. challenge focuses on people tracking, since pedestrians well studied in community, precise detection has...
Existing systems for video-based pose estimation and tracking struggle to perform well on realistic videos with multiple people often fail output body-pose trajectories consistent over time. To address this shortcoming paper introduces PoseTrack which is a new large-scale benchmark human articulated tracking. Our encompasses three tasks focusing i) single-frame multi-person estimation, ii) in videos, iii) establish the benchmark, we collect, annotate release dataset that features labeled...
In this paper, we revisit the joint probabilistic data association (JPDA) technique and propose a novel solution based on recent developments in finding m-best solutions to an integer linear program. The key advantage of approach is that it makes JPDA computationally tractable applications with high target and/or clutter density, such as spot tracking fluorescence microscopy sequences pedestrian surveillance footage. We also show our algorithm embedded simple framework surprisingly...
Abstract Standardized benchmarks have been crucial in pushing the performance of computer vision algorithms, especially since advent deep learning. Although leaderboards should not be over-claimed, they often provide most objective measure and are therefore important guides for research. We present MOTChallenge , a benchmark single-camera Multiple Object Tracking (MOT) launched late 2014, to collect existing new data create framework standardized evaluation multiple object tracking methods....
In this work, we introduce the challenging problem of joint multi-person pose estimation and tracking an unknown number persons in unconstrained videos. Existing methods for images cannot be applied directly to problem, since it also requires solve person association over time addition each person. We therefore propose a novel method that jointly models single formulation. To end, represent body detections video by spatio-temporal graph integer linear program partition into sub-graphs...
The task of tracking multiple targets is often addressed with the so-called tracking-by-detection paradigm, where first step to obtain a set target hypotheses for each frame independently. Tracking can then be regarded as solving two separate, but tightly coupled problems. carry out data association, i.e., determine origin available observations. second problem reconstruct actual trajectories that describe spatio-temporal motion pattern individual target. former inherently discrete problem,...
Tracking-by-detection has proven to be the most successful strategy address task of tracking multiple targets in unconstrained scenarios [e.g. 40, 53, 55]. Traditionally, a set sparse detections, generated preprocessing step, serves as input high-level tracker whose goal is correctly associate these "dots" over time. An obvious short-coming this approach that information available image sequences simply ignored by thresholding weak detection responses and applying non-maximum suppression. We...
When tracking multiple targets in crowded scenarios, modeling mutual exclusion between distinct becomes important at two levels: (1) data association, each target observation should support most one trajectory and be assigned per frame, (2) estimation, trajectories remain spatially separated all times to avoid collisions. Yet, existing trackers often sidestep these constraints. We address this using a mixed discrete-continuous conditional random field (CRF) that explicitly models both types...
Recently, very deep convolutional neural networks (CNNs) have shown outstanding performance in object recognition and also been the first choice for dense prediction problems such as semantic segmentation depth estimation. However, repeated subsampling operations like pooling or convolution striding CNNs lead to a significant decrease initial image resolution. Here, we present RefineNet, generic multi-path refinement network that explicitly exploits all information available along...
The Amazon Robotics Challenge enlisted sixteen teams to each design a pick-and-place robot for autonomous warehousing, addressing development in robotic vision and manipulation. This paper presents the of our custom-built, cost-effective, Cartesian system Cartman, which won first place competition finals by stowing 14 (out 16) picking all 9 items 27 minutes, scoring total 272 points. We highlight experience-centred methodology key aspects that contributed competitiveness. believe these are...
Autonomous robotic manipulation in clutter is challenging. A large variety of objects must be perceived complex scenes, where they are partially occluded and embedded among many distractors, often restricted spaces. To tackle these challenges, we developed a deep-learning approach that combines object detection semantic segmentation. The scenes captured with RGB-D cameras, for which depth fusion method. Employing pretrained features makes learning from small annotated datasets possible. We...
People tracking in crowded real-world scenes is challenging due to frequent and long-term occlusions. Recent methods obtain the image evidence from object (people) detectors, but typically use off-the-shelf detectors treat them as black box components. In this paper we argue that for best performance one should explicitly train people on failure cases of overall tracker instead. To end, first propose a novel joint detector combines state-of-the-art single person with pairs people, which...
Evaluating multi-target tracking based on ground truth data is a surprisingly challenging task. Erroneous or ambiguous annotations, numerous evaluation protocols, and the lack of standardized benchmarks make direct quantitative comparison different approaches rather difficult. The goal this paper to raise awareness common pitfalls related objective evaluation. We investigate influence software, training procedures using several publicly available resources, point out limitations current...
Part handling in warehouse automation is challenging if a large variety of items must be accommodated and are stored unordered piles. To foster research this domain, Amazon holds picking challenges. We present our system which achieved second third place the Picking Challenge 2016 tasks. The challenge required participants to pick list from shelf or stow into shelf. Using two deep-learning approaches for object detection semantic segmentation one item model registration method, localizes...
Standardized benchmarks are crucial for the majority of computer vision applications. Although leaderboards and ranking tables should not be over-claimed, often provide most objective measure performance therefore important guides research. The benchmark Multiple Object Tracking, MOTChallenge, was launched with goal to establish a standardized evaluation multiple object tracking methods. challenge focuses on people tracking, since pedestrians well studied in community, precise detection has...