Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

University of Saskatchewan

$41,000.00

Towards Adaptive and Robust Visual Understanding

EF 2024-25

Résumé
Recent breakthroughs in deep learning have led to remarkable advancements in computer vision, empowering visual understanding models to excel in tasks such as image classification, object detection, semantic segmentation, and video action recognition. However, these models, which rely on full supervision and assume the same data distribution between training data and testing data, suffer substantial performance drops when tested on data from different distributions or domains. For instance, in autonomous driving, a 3D object detection model trained on fully labeled point cloud data from one geographic location may severely underperform when tested on point cloud data from another geographic location. This inability to excel in new domains with distinct data distributions from training data poses a critical hurdle to deploying these models in the real world, hindering their widespread adoption. While a seemingly simple solution could involve collecting labeled training datasets for each new domain and subsequently training the model using supervised learning, this method is neither practical nor financially viable. Hence, my research program seeks to create innovative algorithms that greatly improve visual understanding models' ability to adapt and remain robust in novel domains. In response to this challenge, domain adaptation, which facilitates the adaptation from one domain to another, has emerged as a crucial area of research in computer vision and machine learning. While promising, the existing methods often yield much inferior performance compared to supervised learning in novel domains, limiting their practical utility in real-world applications. In the short term, I aim to pioneer advanced domain adaptation techniques, striving for performance comparable to supervised learning, thus bridging the gap between theoretical progress and practical use. I will explore three specific themes in domain adaptation. Firstly, I will develop novel algorithms that will involve creating innovative gradient agreement strategies during model optimization to enable seamless adaptation from one domain to another and also explore the case when limited labeled data is available in the new domain. Secondly, I will improve domain adaptation by integrating all available modalities. For example, in domain adaptation for video data, visual information is typically utilized, while audio, despite its potential value, is often disregarded. Lastly, I will delve into more practical and realistic settings involving multi-source and multi-target domain adaptation. This proposal will train 2 PhD, 3 MSc, and 5 UG students and will develop a novel set of powerful algorithms to substantially enhance the adaptability and performance of visual understanding models in novel scenarios. It will demonstrate its practical value by showcasing effectiveness with various types of visual data, including images, videos, and point clouds, in a wide range of tasks and applications.
Programme
Discovery Grants Program - Individual
Emplacement
Saskatchewan, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2024-05770
Name nom
Rochan, Mrigank
Department département
Department of Computer Science
Organization id
16
Institution établissement
University of Saskatchewan
Province en
Saskatchewan
Province fr
Saskatchewan
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2024
Award amount
41000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
2603
Research subject group en
Robotics
Research subject group fr
Robotique
Research subject en
Computer vision
Research subject fr
Vision artificielle
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Towards Adaptive and Robust Visual Understanding
Keywords
computer vision; visual understanding; machine learning; deep learning
Application summary
Recent breakthroughs in deep learning have led to remarkable advancements in computer vision, empowering visual understanding models to excel in tasks such as image classification, object detection, semantic segmentation, and video action recognition. However, these models, which rely on full supervision and assume the same data distribution between training data and testing data, suffer substantial performance drops when tested on data from different distributions or domains. For instance, in autonomous driving, a 3D object detection model trained on fully labeled point cloud data from one geographic location may severely underperform when tested on point cloud data from another geographic location. This inability to excel in new domains with distinct data distributions from training data poses a critical hurdle to deploying these models in the real world, hindering their widespread adoption. While a seemingly simple solution could involve collecting labeled training datasets for each new domain and subsequently training the model using supervised learning, this method is neither practical nor financially viable. Hence, my research program seeks to create innovative algorithms that greatly improve visual understanding models' ability to adapt and remain robust in novel domains. In response to this challenge, domain adaptation, which facilitates the adaptation from one domain to another, has emerged as a crucial area of research in computer vision and machine learning. While promising, the existing methods often yield much inferior performance compared to supervised learning in novel domains, limiting their practical utility in real-world applications. In the short term, I aim to pioneer advanced domain adaptation techniques, striving for performance comparable to supervised learning, thus bridging the gap between theoretical progress and practical use. I will explore three specific themes in domain adaptation. Firstly, I will develop novel algorithms that will involve creating innovative gradient agreement strategies during model optimization to enable seamless adaptation from one domain to another and also explore the case when limited labeled data is available in the new domain. Secondly, I will improve domain adaptation by integrating all available modalities. For example, in domain adaptation for video data, visual information is typically utilized, while audio, despite its potential value, is often disregarded. Lastly, I will delve into more practical and realistic settings involving multi-source and multi-target domain adaptation. This proposal will train 2 PhD, 3 MSc, and 5 UG students and will develop a novel set of powerful algorithms to substantially enhance the adaptability and performance of visual understanding models in novel scenarios. It will demonstrate its practical value by showcasing effectiveness with various types of visual data, including images, videos, and point clouds, in a wide range of tasks and applications.