Continual Learning Video Segmentation

Overview

This PhD research addresses the critical challenge of memory-efficient continual learning for video object segmentation in long video sequences. The work develops novel approaches that maintain performance while managing computational constraints.

Research Challenges

• Memory Constraints: Managing limited memory resources for long video processing
• Catastrophic Forgetting: Preventing loss of previously learned knowledge
• Temporal Consistency: Maintaining segmentation consistency across frames
• Real-time Processing: Enabling efficient processing of long video sequences

Technical Contributions

• Memory-Efficient Algorithms: Novel approaches to reduce memory footprint
• Continual Learning Strategies: Techniques to prevent catastrophic forgetting
• Temporal Modeling: Advanced methods for maintaining temporal consistency
• Dataset Development: CLVOS23 dataset for continual learning evaluation

Key Publications

• Nazemi, A. “Continual learning-based Video Object Segmentation.” Doctoral dissertation, University of Waterloo, 2023.
• Nazemi, A., Moustafa, Z., Fieguth, P. “CLVOS23: A long video object segmentation dataset for continual learning.” IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.
• Nazemi, A., Shafiee, M.J., Gharaee, Z. and Fieguth, P. “Memory-Efficient Continual Learning Object Segmentation for Long Videos.” IEEE Access, 2024.

Dataset

CLVOS23: A comprehensive dataset designed specifically for evaluating continual learning approaches in video object segmentation, featuring long video sequences with multiple object categories.

Impact

This research contributes to the field of computer vision by addressing fundamental challenges in video processing, enabling more efficient and practical applications in video analysis, surveillance, and autonomous systems.