Amaya Dharmasiri

I am a first year PhD student in the (Princeton Visual AI lab) advised by Dr. Olga Russakovsky.

I did my undergraduate studies at University of Moratuwa, Sri Lanka, where I completed my bachelors thesis on Learning representations for 3D pointclouds and semantic manipulation of 3D pointcloud objects. The thesis was advised by Dr. Ranga Rodrigo and Dr. Kanchana Thilakarathna.

Previously, I interned as a visiting student researcher at the School of Computer Science, University of Sydney. Later I worked as a research assistant in Mohammad Bin Zayed University of Artificial Intelligence (MBZUAI) in the United Arab Emirates advised by Dr. Salman H Khan, of MBZUAI and Dr. Sadeep Jayasumana of Google Research.

My broader research interests are in vision-language models, multimodal learning, and self-supervised learning.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Developed a point cloud autoencoder architecture to apply controllable transformations to 3D objects via latent space manipulations while preserving perceptual quality of 3D objects and their intended utility in Mixed Reality applications.

Introduced a joint learning objective encapsulating intra-modal correspondence within point cloud modality and cross-modal correspondence between point cloud and 2D image modalities, leveraging contrastive learning. Produced state-of-the-art performance in downstream tasks; 3D object classification, few-shot object classification and 3D object part segmentation.

Proposed a privacy-preserving encoding for 3D point cloud objects with a continuous privilege spectrum for user-controllability of object privacy in mixed-reality applications. Evaluated the performance of the encoding against simulated object reidentification attacks using state-of-the-art point cloud classifiers.

Developed a dynamic categorisation framework for 360-degree video segments based on user viewport trends. This categorization is intended to support efficient caching and resource utilisation in tile-based 360 degree video streaming process.

We investigate different aspects of combining object detections between pairs of stereo images for a given object detector to improve the overall precision and recall. We also deduce a theoretical maximum for this improvement based on the correlation between the stereo pairs.