Yasiru Ranasinghe

SINR: Sparsity Driven Compressed Implicit Neural Representations

Authors: Dhananjaya Jayasundara, Sudarshan Rajagopalan, Yasiru Ranasinghe, Trac D. Tran, and Vishal M. Patel

This work presents a novel compression approach for implicit neural representations (INRs) by encoding their weight space using a sparse high-dimensional dictionary. Unlike prior INR compression methods, SINR avoids transmitting learned dictionaries and remains compatible with existing INR frameworks. It achieves significant storage savings while preserving reconstruction quality across diverse signal types including images, occupancy fields, and NeRFs.

CrowdDiff: Multi-hypothesis Crowd Density Estimation using Diffusion Models

Authors : Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M. Patel

The paper proposes a novel approach to perform crowd counting with multi-hypothesis aggregation using denoising diffusion probabilistic models. The approach outperforms existing state-of-the-art methods for crowd counting.

MonoDiff: Monocular 3D Object Detection and Pose Estimation with Diffusion Models

Authors : Yasiru Ranasinghe, Deepti Hegde, and Vishal M. Patel

The paper proposes using diffusion models to perform monocular 3D object detection and pose estimation. MonoDiff does not require additional modalities to generate intermediate representations to produce box parameters.

Enhancing Crowd Counting with Synthetic Retrieval

Authors: Yasiru Ranasinghe, and Vishal M. Patel

This paper proposes a retrieval-augmented framework for crowd counting that uses synthetic images and text descriptions to enrich visual embeddings. It introduces the use of non-parametric, text-guided knowledge during inference for improved generalization in zero-shot settings. The method achieves state-of-the-art performance on five benchmark datasets without relying on pseudo-labels or manual annotations.

Foundation Model Synergy for Annotator-Free Object Detection

Authors: Yasiru Ranasinghe, Celso M. de Melo, and Vishal M. Patel

Proposes an annotator-free object detection framework using foundational models—SAM and vision-language models (VLMs)—to autonomously generate and label object proposals. The pipeline adapts in real time to novel and domain-specific environments, eliminating the need for manual annotations while achieving strong performance across standard benchmarks.

Eigen-Target Decorrelation for Robust Self-Supervised Learning

Authors: Yasiru Ranasinghe and Vishal M. Patel

Introduces ET-Barlow Twins, a self-supervised learning approach that replaces the identity target with a data-adaptive eigen-target from top-N singular vectors. A student-teacher framework ensures stable training by aligning cross-view correlations. The method improves optimization robustness and generalization over standard Barlow Twins and other SSL baselines.

Zero-Shot Scene Understanding for Automatic Target Recognition Using Large Vision-Language Models

Authors: Yasiru Ranasinghe, Vibashan VS, James Uplinger, Celso De Melo, and Vishal M. Patel

This work presents a robust pipeline for zero-shot automatic target recognition by combining the localization power of open-world detectors with the recognition abilities of large vision-language models (LVLMs). It evaluates multiple LVLMs on military vehicles under challenging conditions and highlights strategies for reliable recognition in novel domains.

Crowd Detection via Point Localization with Diffusion Models

Authors : Yasiru Ranasinghe, and Vishal M. Patel

The paper proposes using diffusion models to localize crowd in images and count the number of people in a scene. The proposed method generates the head locations as a generative task without a separate detector for point proposals.

Authors : Yasiru Ranasinghe, Sanjaya Herath, Kavinga Weerasooriya, Mevan Ekanayake, Roshan Godaliyadda, Parakrama Ekanayake, and Vijitha Herath

The paper process a convolutional autoencoder to realize the nonnegative matrix factorization using deep learning architectures. The proposed methods is applied to extract endmembers and abundances of hyperspectral remote sensing data. The proposed method produced state-of-the-art results on abundance estimation and competitive results in terms of endmember extraction.

Authors : DYL Ranasinghe, HMS Lakmal, HMHK Weerasooriya, EMMB Ekanayake, GMRI Godaliyadda, HMVR Herath, and MPB Ekanayake

The paper proposed an algorithm to determine the availability of surface limestone using hyperspectral satellite imagery of an area. We incorporate traditional image and signal processing, and statistical data analysis techniques in the algorithm. Further, we generated a self-supervisied representation for the hyperspectral signature of limestone in the absence of a groundtruth to improve classification accuracy and spatial continuity of the probability map.

Authors : DYL Ranasinghe , HK Weerasooriya, S Herath, MP Bandara Ekanayake, HMVR Herath, GMRI Godaliyadda, and Terrence Madhujith

The article process a novel non invasive method for food quality analysis, specifically in terms of adulteration, using multispectral images and statistical signal processing and image processing. The proposed method yielded statistically significant results and with significant practical implications.

Authors : EMMB Ekanayake, HMHK Weerasooriya, DYL Ranasinghe , S Herath, B Rathnayake, GMRI Godaliyadda, MPB Ekanayake, and HMVR Herath

The paper proposes a novel algorithm to extract endmembers and abundances of hyperspectral remote sensing data. We introduced a regularizer for the nonnegative matrix factorization which improves independence of endmember statistics. The algorithm illustrated superior performance interms of endmber extraction from hyperspectral data.