Priyank Pathak

I am a third-year Ph.D. student in Computer Vision at the Center for Research in Computer Vision (CRCV), University of Central Florida (UCF), where I am fortunate to be advised by Prof. Yogesh S. Rawat. My research focuses on Low-Resolution Recognition, Person Re-Identification (ReID), and Robustness in Deep Learning. Prior to joining UCF, I earned my Master’s in Computer Science from New York University (NYU) Courant Institute, where I completed my thesis under the guidance of Prof. Rob Fergus. I hold dual Bachelor of Technology (B.Tech) degrees in Electrical Engineering and Computer Science from the Indian Institute of Technology (IIT) Kanpur.

Email: priyank@ucf.edu  /  pp1953@nyu.edu

My research focuses on developing efficient training techniques for computer vision models, with a particular emphasis on understanding the adverse effect of noise and enhancing the robustness of deep learning architectures in real-world scenarios. As a Ph.D. student, my work investigates how various types of noise—such as low-resolution and environmental distortions—impact the performance of deep learning models, including transformers and vision-language models (VLMs).   Previously, my Master's thesis centered on Person Re-Identification (ReID) in surveillance settings, a domain inherently affected by low-resolution and noisy visual data.

We propose a lightweight, annotation-free proxy for mitigating appearance bias in ReID models, when clothing annotations aren't available. We propose Colors See, Colors Ignore (CSCI), an RGB-only method that leverages color information directly from raw images or video frames. CSCI efficiently captures color-related appearance bias (‘Color See’) while disentangling it from identity-relevant ReID features (‘Color Ignore’).

We introduce LR0.FM, a benchmark evaluating the impact of low resolution on the zero-shot classification performance of VLMs. We propose a novel metric, Weighted Aggregated Robustness, to address the limitations of existing metrics and LR-TK0, to enhance the robustness of models without compromising their pre-trained weights. We demonstrate the effectiveness of LR-TK0 for robustness against low-resolution across several datasets and its generalization capability across backbones and other approaches.

One of the first works to solve Person Reid exploring multiple techniques to improve accuracy on the Video ReID model.

We critically examine the SOTA Zero-Shot Object Detectors' vulnerability to resolution degradation (`pixelation'). Two key takeaways are 1) The choice of backbone alone can substantially enhance robustness, while prompting (language) plays a minimal role. 2) Robustness is also a function of the dataset/domain of images at inference. A certain type of image are not affected by pixelation.

We introduce Robustness against Low-Quality (RLQ) in clothes changing real-world ReID to make the model robust against low-quality artifacts like pixelation, out-of-focus blur, and motion blur.

I sometimes release implementations for others' papers.

"Aloo has no plaace in Biryani" - Every Desi Foodie

Built upon Jon Barron's template, and modified upon Rohit Gupta