Logo: University of Southern California

Events Calendar



Select a calendar:



Filter April Events by Event Type:



University Calendar
Events for April

  • PhD Thesis Defense - Hikaru Ibayashi

    Tue, Apr 11, 2023 @ 09:00 AM - 10:30 AM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Hikaru Ibayashi

    Title:
    Sharpness Analysis of Neural-networks for Physics Simulation

    Committee members:
    Prof. Aiichiro Nakano (Chair), Prof. Yan Liu, Prof. Paulo Branicio (Department of Chemical Engineering and Materials Science)

    Abstract:
    Deep learning has attracted significant attention in recent years due to its remarkable achievements in various applications. However, building effective deep neural networks requires making crucial design choices such as the network architecture, regularization, optimization, and hyperparameter tuning.
    In this dissertation, we focus on the concept of ``sharpness'' of neural networks,
    which refers to neural networks' sensitivity against perturbation on weight parameters. We argue that sharpness is not only a theoretical notion but also has practical use cases that can lead to better generalization and robustness of neural models.

    A major theoretical challenge of defining and measuring sharpness is its scale-sensitivity, i.e., the fact that sharpness can change to the scale transformation of neural networks. In this thesis, we propose a novel definition of sharpness that overcomes this limitation, with provable scale-invariance and extensive empirical validation. By analyzing the relationship between sharpness and model performance, We show how my definition can provide a more objective and accurate characterization of sharpness.

    Another open question in the sharpness analysis is how training algorithms for machine learning models regularize sharpness. In this dissertation, we answer this question by showing that existing training algorithm methods regularize sharpness through what can be called "escaping" behavior, where the optimization process avoids sharp regions in the parameter space. This new explanation demystifies the connection between sharpness and training algorithms, paving the way for more effective and principled approaches to machine learning.

    Finally, we demonstrate the practical benefits of sharpness regularization for physics simulations. We show that neural networks with small sharpness achieve high-fidelity fluid simulation and molecular dynamics. These findings include the significant implication that sharpness is not just a mathematical notion but also a practical tool for building physics-informed neural networks.

    Location: Seaver Science Library (SSL) - 104

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • ShowCAIS 2023

    Fri, Apr 14, 2023 @ 09:00 AM - 05:00 PM

    Thomas Lord Department of Computer Science, USC Viterbi School of Engineering

    University Calendar


    The USC Center for AI in Society is hosting an all-day, in-person symposium on Friday, April 14th on the USC campus. This event will highlight the work of students and faculty using AI for good, and will include lunch and refreshments.

    Registration: https://sites.google.com/usc.edu/showcais2023/registration?authuser=0

    PLEASE REACH OUT TO THE SHOWCAIS ORGANIZING COMMITTEE WITH ANY QUESTIONS: USCCAIS@USC.EDU

    Location: Ronald Tutor Hall (RTH) 526

    Audiences: Everyone Is Invited

    Contact: Caitlin Dawson

    Event Link: https://sites.google.com/usc.edu/showcais2023/home?authuser=0

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PHD Thesis Proposal - Julie Jiang

    Mon, Apr 17, 2023 @ 01:00 PM - 02:30 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PHD Thesis Proposal: Julie Jiang

    Committee: Emilio Ferrara (Chair), Barath Raghavan, Su Jung Kim, Jesse Thomason, Kristina Lerman

    Title: Socially-infused Content Mining of Online Human Behavior

    Abstract:
    The vast amount of data generated by human behavior online provides valuable insight into how people interact with one another and with digital environments. However, mining this data can be time-consuming and computationally intensive. This dissertation proposes a unified language and network model that leverages the concept of homophily to efficiently analyze large-scale human behavior. By identifying patterns in network interactions and linguistic styles, this model can characterize political polarization, detect hateful and toxic users, and quantify users based on their moral foundation leanings. The findings demonstrate how seemingly simple patterns in online behavior can offer a deeper understanding of human behavior in digital environments. I apply this model to a range of real-world problems, including characterizing political polarization, understanding social influence on networks of hateful users, and contextualizing user behavior based on their moral foundation leanings. The findings demonstrate how seemingly simple patterns in online behavior can offer a deeper understanding of human behavior in digital environments.


    Location: https://usc.zoom.us/j/96953099505?pwd=MDhJVFFSbDhuNnBWNm9JZjRFRUVjZz09

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Event Link: https://usc.zoom.us/j/96953099505?pwd=MDhJVFFSbDhuNnBWNm9JZjRFRUVjZz09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Haidong Zhu

    Mon, Apr 17, 2023 @ 02:00 PM - 03:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    Title: Body Shape Reconstruction for Video-based Person Identification

    Committee: Ram Nevatia (Chair), Ulrich Neumann, Mohammad Soleymani, Stefanos Nikolaidis, Antonio Ortega

    Date: Monday April 17, 2pm PST

    Abstract: The recognition of individuals in videos is a crucial task in connecting video clips of the same person captured by multiple non-overlapping cameras. While image-based person identification mainly relies on the person's appearance, video-based identification can exploit additional external information, such as walking poses and general body shape, leading to less biased identification results. However, changes in body shape due to clothing and camera views can negatively impact such biometrics. To address this limitation, we propose reconstructing 3-D human body shapes across frames to provide consistency and invariance to view changes, improving identification accuracy and robustness. Existing 3-D body shape reconstruction methods typically focus on frame-by-frame reconstruction and fail to leverage the consistency between frames. In this proposal, we aim to enhance body shape reconstruction and representation of objects and body. Additionally, we will extend these methods to reconstruct body shapes in videos and utilize this information to aid person identification. Our proposed approach leverages temporal information between frames and utilizes body shape information for identification assistance, with the goal of improving identification accuracy.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/94346121440?pwd=S3JSVVRpTFc5WFB5THEvWE9TTEhSQT09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Defense - Naghmeh Zamani

    Thu, Apr 20, 2023 @ 11:00 AM - 01:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Naghmeh Zamani

    Title: Perception and Haptic Interface Design for Rendering Hardness and Stiffness

    Committee: Heather Culbertson, Jernej Barbic, Somil Bansal

    Abstract: In this talk, I will discuss the challenge of accurately rendering the sensations of hardness and stiffness in haptic applications, which is a critical problem for applications such as medical simulation that require accurate virtual hardness and stiffness replication. The first part of the talk will present a set of experiments to investigate human tactile perception sensitivity in tool-mediated systems. The second part will explore a new method for rendering hard objects using an encountered-type haptic display and augmented reality. The talk will evaluate how changing the hardness of the end-effector affects the user's perception of the interaction and proposes a dynamic end-effector for a more accurate and realistic simulation of hardness and stiffness. Furthermore, I will discuss the investigation of the underlying events on the skin during the interaction between a bare finger and the environment. The results suggest that the spectral content of vibration feedback is important mechanical information for surface hardness discrimination and natural material identification. The talk will provide insights and solutions to improve the accuracy and realism of haptic simulations for applications that require the perception of hardness and stiffness in virtual objects

    Location: Charles Lee Powell Hall (PHE) - 325

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Defense - Brendan Avent

    Fri, Apr 21, 2023 @ 04:00 PM - 06:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Brendan Avent

    Title: Practice-Inspired Models and Mechanisms for Differential Privacy

    Committee Members: Aleksandra Korolova (chair), Salman Avestimehr (Department of Electrical and Computer Engineering), Leana Golubchik, David Kempe, Cyrus Shahabi


    Abstract: Now more than ever, organizations such as companies, governments, and researchers must collect and analyze people's sensitive data to drive decisions and fuel innovation. Differential privacy has become the gold standard for data privacy in computer science literature, particularly for privacy-preserving data analysis and machine learning. Significant research effort has been devoted to designing and theoretically analyzing mechanisms that satisfy differential privacy. However, far less research has studied the pragmatic considerations of differential privacy, i.e., how its trust models and mechanisms can be adapted and applied for real-world uses.

    I focus on making differential privacy useful for real-world applications by removing barriers that hinder its adoption in practice. In the first part, I address the utility gap between the more and less desirable trust models of differential privacy by defining and analyzing a new hybrid trust model. In the second part, I address the lack of tools for analyzing the utility of complex differentially private mechanisms by developing a new method for quantifying such mechanisms privacy--utility trade-offs. Finally, I show how to improve the utility of DP mechanisms that answer statistical queries on a large scale. In the classic setting where all queries are provided to the mechanism in advance, I detail how we extend the state-of-the-art differentially private mechanism for answering marginal queries to a more general, flexible query class. I then define a new setting where our extended mechanism is only provided partial knowledge of which queries will be posed. Analyzing the mechanism in both settings, I show that it answers a massive number of queries both efficiently and effectively.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/95544425859?pwd=Wk82dTBEQkhyMDFxeGtqS2VqK0h5UT09 ;Meeting ID: 955 4442 5859 ;Passcode: 645176

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PHD Thesis Defense - Christopher Denniston

    Mon, Apr 24, 2023 @ 02:00 PM - 04:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PHD Thesis Defense - Christopher Denniston

    Title: Active Sensing In Robotic Deployment

    Committee Members: Prof. Gaurav S. Sukhatme (Chair), Prof. Jesse Thomason, Prof. David A. Caron


    Date/Time: April 24th, 2-4pm

    Location: RTH 306 or on Zoom https://usc.zoom.us/j/2869134593

    Abstract:

    Robots have the potential to greatly increase our ability to measure complex environmental phenomena, such as harmful algae blooms, which can harm humans and animals alike in drinking water. Such phenomena require study and measurement at a scale that is beyond what can be accomplished by robots that plan to completely cover the area. Despite this, many sensing robots still are deployed with non-active behaviors, such as fixed back-and-forth patterns. The lack of deployment of active sensing systems in practice is due to difficulties with problems encountered in the real world. We identify and address solutions for three main issues which plague complex real robotic active sensing deployments.

    First, active sensing systems are difficult to use, with complex deployment-time decisions that affect the efficiency of sensing. We describe systems that eschew these decisions, allowing for efficient and automatic deployment. We find that these systems provide a non-technical deployment procedure and outperform hand-tuned behaviors.

    Second, active sensing robots tend to perform a survey that maximizes some general goal and requires the user to interpret the collected data. We propose a system that, instead, plans for the specific user task of collecting physical samples at limited, unknown locations. We demonstrate that planning for this specific task while sensing allows for more efficiency in the active sensing behavior.

    Finally, existing models for active sensing do not accurately model the interaction of the sensed signal and obstacles in the environment. We propose two novel modeling techniques which allow active sensing of signals which have complex interactions with obstacles, such as electromagnetic waves. Both outperform traditional modeling techniques and enable scalable active sensing to a large number of measurements on a real robot. Additionally, we find that they allow the robot to actively place signal-emitting devices while sensing the signals from these placed devices.

    Location: Ronald Tutor Hall of Engineering (RTH) - 306

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Defense - Yannan Li

    Tue, Apr 25, 2023 @ 10:00 AM - 12:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Yannan Li

    Title: Formal Analysis of the Data Poisoning Robustness of K-Nearest Neighbors

    Committee members(Lexicographic order): Pierluigi Nuzzo, Mukund Raghothaman, Chao Wang (chair)

    Abstract: Data poisoning, which aims to corrupt a machine learning model and change its inference results by changing data elements in its training set, poses a significant threat to machine learning based software systems. However, formally certifying data poisoning robustness is a challenging task. I designed and implemented a set of formal methods for deciding, both efficiently and accurately, the data-poisoning robustness of the k-nearest neighbors (KNN) algorithm, which is a widely-used supervised machine learning technique. First, I developed a method for certifying the data-poisoning robustness of KNN by soundly overapproximating both the learning and inference phases of the KNN algorithm. Second, I developed a method for falsifying data-poisoning robustness, by quickly detecting the truly-non-robust cases using search space pruning and sampling. Finally, I extended these methods to other attack models and fairness certification, thus allowing for a more comprehensive analysis of the robustness of KNN.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/94891715635?pwd=SFI5VFBtMndhN3BORk5GSjRyS2IzQT09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Shihan Lu

    Tue, Apr 25, 2023 @ 11:00 AM - 01:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    Title: Analysis, Synthesis, and Perception of Multimodal Feedback for Humans and Robots

    Committee: Heather Culbertson (Chair), Stefanos Nikolaidis, Gaurav Sukhatme, Jernej Barbic, Shrikanth Narayanan

    Date: Tuesday, April 25, 11 am - 1 pm PST

    Abstract: Multimodal feedback, including haptic and auditory feedback, is often overlooked in interactive and contact-rich scenarios in the studies with both humans and robots, such as writing on the back of an envelope with a pen or grasping a block in a Jenga game. In this work, I focus on three perspectives related to the multimodal feedback in interactions: (1) Analysis â“ how to extract useful and interpretable features from multimodal feedback; (2) Synthesis â“ how to simulate realistic virtual feedback; and (3) Perception â“ how humans and robots respond to the feedback. I explore these perspectives through tasks of sound modeling, haptic texture design, large-scale texture classification, and state-aware robot manipulation. With these tasks, the objective is to enhance the interactive experience in virtual reality, improve the understanding of crossmodal relationships, and complement visual and tactile sensing in challenging robot manipulation tasks.

    Location: Ronald Tutor Hall of Engineering (RTH) - 306

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/93301568398?pwd=RjYxRTRSU1hrbktzN0wweWZxT0JDQT09 ; Meeting ID: 933 0156 8398 ; Passcode: 311777

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PHD Thesis Proposal - Woojeong Jin

    Tue, Apr 25, 2023 @ 12:00 PM - 02:00 PM

    Thomas Lord Department of Computer Science

    University Calendar



    PHD Thesis Proposal - Woojeong Jin

    Title: Towards a Better Reasoner on Visual Information

    Humans acquire knowledge by processing visual information through observation and imagination, which expands our reasoning capability about the physical world we encounter every day. Despite significant progress in solving AI problems, current state-of-the-art models in natural language processing (NLP) and computer vision (CV) have limitations in terms of reasoning and generalization, particularly with complex reasoning on visual information and generalizing to unseen vision-language tasks. This thesis proposal aims to address these shortcomings by presenting a series of works that enable smaller vision-language (VL) models to generalize to new tasks, improve language models by incorporating visual information, and evaluate language models by assessing their ability to reason about the physical world through text.

    https://usc.zoom.us/j/98941948220

    12 pm on 4/25

    Committee Members: Xiang Ren, Ram Nevatia, Jesse Thomason, Robin Jia, Emilio Ferrara.

    Location: https://usc.zoom.us/j/98941948220

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Event Link: https://usc.zoom.us/j/98941948220

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PHD Thesis Proposal - Peifeng Wang

    Tue, Apr 25, 2023 @ 03:00 PM - 04:30 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PHD Thesis Proposal - Peifeng Wang

    Title: Building Small-Scale but Advanced Language Reasoners

    Abstract:
    The entanglement of multiple language capabilities within large language models requires expensive scaling to work effectively. I argue that a disassociation of these capabilities from core language skills can enable the creation of smaller, more accessible language models. Additionally, this disassociation will facilitate the development of language models with enhanced reasoning abilities.

    This thesis proposal presents three techniques to build small language models with advanced reasoning capabilities. First, I introduce an Imagine&Verbalize framework for generative commonsense reasoning, which decomposes a complex generation task into easier sub-tasks and learns from a diverse set of indirect supervision from multiple domains. Second, I present a knowledge-transferring pipeline which prompts large language models to rationalize for an open-domain question and then trains small language models to answer consistently. Third, I discuss augmenting small LMs with a working memory for coherent language reasoning by tracking the states of the described world.



    Venue: zoom at https://usc.zoom.us/j/97850702935?pwd=ekJ0K1RMM045Tk1EQUV1OUEvOE5iQT09

    Date and time: 3:00pm-4:30pm on April 25th

    Committee Members: Xiang Ren (chair), Filip Ilievski, Swabha Swayamdipta, Ram Nevatia, Emilio Ferrara

    Location: https://usc.zoom.us/j/97850702935?pwd=ekJ0K1RMM045Tk1EQUV1OUEvOE5iQT09

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Event Link: https://usc.zoom.us/j/97850702935?pwd=ekJ0K1RMM045Tk1EQUV1OUEvOE5iQT09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Iordanis Fostiropoulos

    Thu, Apr 27, 2023 @ 10:00 AM - 12:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Proposal - Iordanis Fostiropoulos

    Committee: L. Itti (Chair), M. Soleymani, S. Nikolaidis, N. Schweighofer (Outside Member)

    Title: Towards Learning Generalizable Representations

    Abstract: Current work in Machine Learning (ML) research lack systematic tools and methods for evaluating the performance of a ML model on the ability to generalize beyond the train set; where the current accepted practice is on the evaluation of the loss on a test set. Work in ML for defining generalization is abstract and based on anthropocentric measures[65]. While practical metrics in evaluating generalization are poor indicators where there are trade-offs between the metric (such as loss) and the performance of the Deep Neural Network (DNN) to Out-of Distribution examples, such as robustness-accuracy trade-off or hallucinations of transformer models. While algorithmic solutions are often in the form of paradigm shifts that are ad-hoc and domain specific with a lack of consensus in literature. Our work focus on generalization as it pertains on evaluating and improving current ML systems, as opposed to proposing a paradigm shift, where we address three evaluation settings of generalization. First, the generalization of a DNN to learn generalizable representations useful beyond the task it was trained on. Second, the generalization of the learning hyper parameters used to fit a DNN; a meta-model. Third, the learning algorithm generalization, where we evaluate generalization in the context of Continual Learning. We present our work on the analysis and theoretical findings on the short-comings of generalization and provide practical solutions that both confirm and can in-part address the issue. We motivate that the problem of generalization extend well beyond the three areas our work addresses where improvements in algorithms, tools, and methods are required. Finally, based on our empirical observations we discuss several future directions for improving generalization in ML systems.

    Location: Henry Salvatori Computer Science Center (SAL) - 322

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Dissertation Defense - Lauren Klein

    Thu, Apr 27, 2023 @ 01:00 PM - 03:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Dissertation Defense - Lauren Klein

    Committee Members: Maja Mataric (chair), Pat Levitt, Shrikanth Narayanan, Mohammad Soleymani, and Jesse Thomason

    Title: Modeling Dyadic Synchrony with Heterogeneous Data: Validation in Infant Mother and Infant Robot Interaction

    Abstract: Our health and wellbeing are intricately tied to the dynamics of our social interactions, or social synchrony. The key components of social synchrony during embodied interactions are temporal behavior adaptation, joint attention, and shared affective states. To create comprehensive representations of nuanced social interactions, computational models of social synchrony must account for each of these components.

    The goal of this dissertation is to develop and evaluate approaches for modeling social synchrony during embodied dyadic interactions. We present computational models of social synchrony in two contexts. First, we explore human to human social interactions, where attention and affective states must be inferred through behavioral observations. During embodied interactions, social partners communicate using a diverse range of behaviors, therefore, this work develops approaches for modeling temporal behavior adaptation using heterogeneous data, or data representing multiple behavior types. Next, we explore social synchrony in the context of human to robot interaction. Robots must be equipped with perception modules to establish joint attention and shared affective states based on information about their partners behaviors. To address this need, we develop and evaluate models for attention and affective state recognition. Given the central role of communication in cognitive and social development, this dissertation focuses on interactions that occur during infancy and early childhood. Specifically, we develop and evaluate our approaches using recordings of infant to mother, infant to robot, and child to robot interactions.

    The work presented in this dissertation for evaluating and supporting social synchrony enables new opportunities to study the relationships between individual behaviors, joint interaction states, and developmental and health outcomes.

    Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Adriana Sejfia

    Thu, Apr 27, 2023 @ 01:00 PM - 03:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Proposal - Adriana Sejfia

    Committee Members: Nenad Medvidovic (chair), Chao Wang, William Halfond, Mukund Raghothaman, Sandeep Gupta, and Jyotirmoy Deshmukh

    Title: Systematic Improvement of Deep Learning Based Vulnerability Detection

    Abstract: Deep learning based techniques have gained traction in software vulnerability detection. However, the performance of these techniques in data drawn from distributions other than the ones the models have been explicitly trained on has been shown to vary a lot. In this talk, I will present our study on four limitations of the current deep learning based vulnerability detectors and the datasets they use along with solutions we propose to address these limitations

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/97573523067?pwd=aW94cUlkM3IwZmk5L3E2a1ZTTG9SUT09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Defense - Cho-Ying Wu

    Fri, Apr 28, 2023 @ 11:30 AM - 12:30 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Cho-Ying Wu

    Committee Members: Ulrich Neumann (chair), Laurent Itti, Andrew Nealen, C.C. Jay Kuo

    Title: Meta Learning for Single Image Depth Prediction

    Abstract: Predicting geometry from images is a fundamental and popular task in computer vision and has multiple applications. For example, predicting ranges from ego view images can help robots navigate through indoor spaces and avoid collisions. Additional to physical applications, one can synthesize novel views from single images with the help of depth by warping pixels to different camera positions. Further, one can fuse depth estimation from multiple views and create a complete 3D environment for AR VR uses.

    In the dissertation, we aim to discover a better learning strategy, meta learning, to learn a higher level representation. The learned representation more accurately characterizes the depth domain. Our presented meta learning approach attains better performance without involving extra data or pretrained models but directly focuses on learning schedules. Then, we closely evaluate the generalizability on our collected Campus Data and demonstrate meta learning's ability in sub, single, multi dataset levels.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/9340884176

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Defense - Gozde Sahin

    Fri, Apr 28, 2023 @ 02:00 PM - 04:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Thesis Defense - Gozde Sahin

    Title: Towards More Occlusion-Robust Deep Visual Object Tracking

    Committee Members: Prof. Laurent Itti (chair), Prof. Ulrich Neumann, Prof. Keith Jenkins

    Abstract: Visual object tracking (VOT) is considered as one of the principal challenges in computer vision, where a target given in the first frame is tracked in the rest of the video. Major challenges in VOT include factors such as rotations, deformations, illumination changes, and occlusions. With the widespread use of deep learning models with strong representative power, trackers have evolved to better handle the changes in the targets appearance due to factors like rotations and deformations. Meanwhile, robustness to occlusions has not been as widely studied for deep trackers and occlusion representation in VOT datasets has stayed low over the years.

    In this work, we focus on occlusions in deep visual object tracking and examine whether realistic occlusion data and annotations can help with development and evaluation of more occlusion-robust trackers. First, we propose a multi-task occlusion learning framework to show how much occlusion labels in current datasets can help improve tracker performance in occluded frames. We discover that lack of representation in VOT datasets creates a barrier for developing and evaluating trackers that focus on occlusions. To address occlusions in visual tracking more directly, we create a large video benchmark for visual object tracking: The Heavy Occlusions in Object Tracking (HOOT) Benchmark. HOOT is specifically tailored for evaluation, analysis and development of occlusion-robust trackers with its extensive occlusion annotations. Finally, using the annotations in HOOT, we examine the effect of occlusions on template update and propose an occlusion-aware template update framework that improves the tracker performance under heavy occlusions.

    Location: Hedco Neurosciences Building (HNB) - 100

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File