Arka Mitra

I recently graduated my masters programme in the Department of Information Technology and Electrical Engineering at ETH Zurich, specializing in machine learning and signal processing. I am interested in computer vision, deep learning, generative AI, and image processing. Most of my research interest is about scene understanding and having a higher level understanding of the physical world. I currently hold a valid work B permit in Switzerland and am actively seeking for a job.

I am also responsible for the computer vision group at NomadZ, the robosoccer team at ETH Zurich where we teach robots to play soccer autonomously.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github  /  LinkedIn

News

• Optimizing Long-Term Player Tracking and Identification in NAO Robot Soccer by fusing Game-state and External Video was accepted in ICRA 2023.
• NomadZ visisted Hamburg for GORE 2023.
• Recieved grants from NCCR Automation and KIM for NomadZ.
• Causality Detection using Multiple Annotation Decisions was accepted to EMNLP 2022.
• One of the finalists at HackZurich 2022.
• NomadZ visited Bangkok for RoboCup 2022.
• Helped organize Medical Imaging using Deep Learning conference 2022.
• Joined the computer vision group of NomadZ.
• Joined ETHZ for my masters.
• Joined Microsoft India as a Data and Applied Scientist.
• Graduated from IIT Kharagpur.

Publications

I'm interested in computer vision, deep learning, image processing and natural language processing. Some papers are listed below.

We propose an approach for long term tracking of similar looking Nao Robots using quadratic optimization which combines the robot's internal states and video from non-calibrated camera.

We introduce a dataset consisting of both synthetic and real-world referee signals. Additionally we show the amount of data required to train an end-to-end model using real data if already trained on synthetic data.

We propose an approach for long term tracking of similar looking Nao Robots using quadratic optimization which combines the robot's internal states and video from non-calibrated camera.

We propose a loss which considers the decision from multiple different annotators for causality detection.

We create a model for hate speech and offensive content detection in multiple languages.

We detect early-stage Rheumatoid Arthritis using meta-heuristic search, combinig it with fuzzy logic to provide explainable feture importance.

We show a comparison of classical and deep learning based methods for detecting COVID-19 from CT images.

We evaluated different deep learning models and techniques for tracking drones in videos. It was done as a part of my internship at Honeywell.

We combine machine learning methods with meta-heuristic optimization to speed up the convergence of stable cluster structures.

We demonstrate that embeddings from large language models can be used to detect causality in a sentence. Additionally, we show that models trained on next sentence prediction are better in prediciting causality.

We explain the predictions of deep learning models on chest xrays. The results were validated by a radiologist.

We incorporate domain knowledge into meta-heuristic algorithms to improve the convergence of the algorithms.

We find the most stable structures of carbon clusters using a meta-heuristic algorithm in combination with Gaussian07 software.

Blogposts

I keep track of some of the machine learning papers.