| Title | : | Efficient Semantic Segmentation Approaches for Challenging Scenarios: Adverse Lighting and Sparse Data |
| Speaker | : | Ms. Meena Padnekar S (IITM) |
| Details | : | Wed, 19 Mar, 2025 12:00 PM @ SSB 334 |
| Abstract: | : | Image segmentation is the task of assigning semantic categories to each pixel in the image. The supervised learning paradigm of segmentation has seen considerable advances in recent years, producing models with good generalization capabilities. However, the major factors influencing the generalization capability of these models are the quantity and quality of training data. Acquiring large amounts of labeled quality data is a daunting task, primarily due to the sheer amount of time and effort required to acquire the images with their annotations. These limitations of the supervised learning paradigm have triggered advances in unsupervised or self-supervised learning methods, which perform the same task without using any target labels for training, and few-shot methods that utilize very few samples.
Unsupervised Semantic Segmentation (USS) groups pixels without labeled data, but existing methods relying solely on RGB images struggle to perform well in poor lighting conditions. In this seminar, we introduce DARTS, a novel multimodal USS framework that combines RGB and thermal modalities to enhance segmentation robustness under challenging visibility conditions such as night-time images, artificial light sources, and glare. DARTS effectively captures complementary features and improves representation learning by leveraging a self-supervised Vision Transformer (ViT) and a multimodal fusion module with feature correlation loss. Our approach demonstrates superior performance across multiple benchmarks, particularly in low-light and high-glare environments, and can be incorporated into the existing USS methods for efficient performance.
Join us to explore how DARTS advances USS beyond traditional unimodal approaches and its potential impact on real-world applications.
Web Conference: meet.google.com/tgr-nuxe-cyh |
