Nyle Siddiqui

I'm a 5th-year Ph.D student at the Center for Research in Computer Vision at the University of Central Florida advised by Dr. Mubarak Shah. I am also a NSF GRFP 2024 Honorable Mention.

Email  /  CV  /  Scholar  /  Github  /  LinkedIn

Research

My research broadly spans the field of computer vision, with specific interests including generative AI, diffusion models, action recognition, person recognition, and representation learning. You can check out my selected papers below, with important papers highlighted.

We introduce a geometry-aware knowledge distillation framework for efficiently training a large-scale state space model for dense 3D reconstruction, achieving comparable reconstruction accuracy with substantially faster inference.

We investigate the use of SSMs for continuous, real-time EEG brain signal decoding.

We introduce VRR-QA, a benchmark for visual relational reasoning in videos that tests whether models can infer relationships, actions, and contextual cues that are not explicitly shown. The benchmark contains 1K expert-annotated question-answer pairs across 1K creative video clips and reveals a substantial gap between current VideoQA models and human performance on implicit reasoning.

We propose a novel transformer decoder-based architecture in tandem with two supervised contrastive losses for multi-view action recognition. By disentangling the view-relevant features from action-relevant features, we enable our model to learn action features that are robust to change in viewpoints. We show that changes in viewpoint impart perturbations on learned action features, and thus, disentangling these perturbations improves overall action recognition performance.

We propose a generative data expansion framework via diffusion for clothes-changing person Re-ID, which leverages pre-trained diffusion models and large language models to accurately generate images of individuals with different clothing attires. We address the challenges faced by CC-ReID models due to the limited clothing diversity in current CC-ReID datasets by genereating additional synthetic data that increases clothing diversity while preserving important personal features in the generated images. We also introduce two novel CC-ReID training strategies: progressive learning and test-time prediction refinement. Notably, training certain models with data generated by DLCR on the PRCC dataset resulted in improvements of up to 11.3% improvement in top-1 accuracy, with additonal enhanced performance on out-of-distribution test data.

Currently under review. We instill VideoMamba with spatio-temporal flexibility and shows it performs better on a variety of action recognition tasks.

Internship Experience

Conducted research on video object detection and grounding using state space models for Ring cameras. Paper submitted to CVPR 2027.

Conducted novel research applying vision-language models to polarimetric data for facial anti-spoofing.

Awards / Recognitions

Professional Reviewing Experience

Mentor in NSF-REU