Akash Kumar

I am a fifth year PhD student at Center for Research in Computer Vision (CRCV), University of Central Florida (UCF), under the supervision of Prof. Yogesh Singh Rawat.

I have a broad interest in deep learning and computer vision. My current research mainly focuses on data-efficient approaches for dense video tasks.

Looking for research intern/full-time positions (Summer'25)! Feel free to drop me an email.

Email  /  Google Scholar  /  Github  /  LinkedIn  /  CV

Updates

  Jan'25: First author paper accepted at ICLR 2025 💥
  Jan'25: Selected for Doctoral consortium at IEEE/CVF WACV 2025 💥
  Dec'24: First author paper accepted at AAAI 2025 💥
  May'24: Started internship at Amazon, Palo Alto, CA
  Dec'23: A paper accepted to AAAI 2024 💥
  Oct'23: A First author paper accepted at NeurIPS Self-Superivsed Workshop 2024
  Mar'23: A paper accepted to CVPR Workshops 2022
  Mar'22: A First author paper "E2E-SSL" accepted to CVPR 2022 💥

Towards Open-vocabulary video object understanding.

Below is a selected list of my works (in chronological order), representative papers are highlighted.

Adaptation of VLMs leveraging Large Language Models (LLMs) via spatio-temporal composite relationship.

Improved VLMs grounding capabilities via action composition and complex spatio-temporal scene understanding.

Developed first vision language models (VLMs) for dense multimodal video detection task without any labels. Devised context aware and self-paced progressive scene learning approach.

Learning from mistakes on labelled set and transfer that learning to pseudo labels from unlabeled set to enhance spatio-temporal localization. Class-agnostic spatio-temporal refinement module and temporal coherency constraint for better spatio-temporal localization.

High-pass filtering for enhanced pseudo labels to improvise spatio-temporal localization. Simple sample augmentation strategy for informative sample selection.

First exhaustive study on impact of pre-training in self-supervised learning for videos. Proposed a simple knowledge distillation approach outperforming previous works with 90% less videos.

First end-to-end semi-supervised approach for video action detection task. Short-term and long-term smoothness constraints to exploit spatio-temporal coherency.

Developed new spatio-temporal surveillance based dataset for real-world challenges.

Proposed real-time online action detection system for open-world surviellance videos.

Feel free to steal this website's source code. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead. Also, consider using Leonid Keselman's Jekyll fork of this page.