Arkadip Maitra

I am a machine learning engineer at PrediQt, Kolkata. I train and work with language and vision models that are used as virtual sales assistants for e-commerce applications. I also work as a researcher at Department of Computer Science, RKMVERI, Belur. I work on sign language recognition and generation tasks.

 |  Email  |   |  LinkedIn  | 
 |  CV  |  Github  | 
 |  Google Scholar  | 

I am a Sr. AI/ML Engineer at PrediQt, Kolkata, where I work with multinational companies on development and use of language and vision models. I also work as a researcher at Department of Computer Science, RKMVERI, Belur. I work on sign language recognition and generation with Suvajit Patra, under the guidance of Prof. Soumitra Samanta I have worked on self supervised learning and writer identification with Siladittya Manna, Senior Research Assistant at the Department of Computer Science, Hong Kong Baptist University and Prof. Umapada Pal and Prof. Saumik Bhattacharya. I have also worked on earthquake early warning system under the guidance of Prof. Amlan Chakrabarti, Professor and Director, A.K. Choudhury School of Information Technology, University of Calcutta. I did my B.Sc in Computer Science from RKMVCC and M.Sc in Computer Science from RKMVERI.

Automatic Sign Language (SL) recognition is an important task in the computer vision community. To build a robust SL recognition system, we need a considerable amount of data which is lacking particularly in Indian sign language (ISL). In this paper, we propose a large-scale isolated ISL dataset and a novel SL recognition model based on skeleton graph structure. The dataset covers 2,002 daily used common words in the deaf community recorded by 20 (10 male and 10 female) deaf adult signers (contains 40033 videos). We propose a SL recognition model namely Hierarchical Windowed Graph Attention Network (HWGAT) by utilizing the human upper body skeleton graph structure. The HWGAT tries to capture distinctive motions by giving attention to different body parts induced by the human skeleton graph structure. The utility of the proposed dataset and the usefulness of our model are evaluated through extensive experiments. We pre-trained the proposed model on the proposed dataset and fine-tuned it across different sign language datasets further boosting the performance of 1.10, 0.46, 0.78, and 6.84 percentage points on INCLUDE, LSA64, AUTSL and WLASL respectively compared to the existing stateof-the-art skeleton-based models. The proposed dataset and the model implementation will be available at https://cs.rkmvu.ac.in/~isl.

Self-supervised learning has developed rapidly over the last decade and has been applied in many areas of computer vision. Decorrelation-based self-supervised pretraining has shown great promise among non-contrastive algorithms, yielding performance at par with supervised and contrastive self-supervised baselines. In this work, we explore the decorrelation-based paradigm of self- supervised learning and apply the same to learning disentangled stroke features for writer identification. Here we propose a modified formulation of the decorrelation-based framework named SWIS which was proposed for signature verification by standardizing the features along each dimension on top of the existing framework. We show that the proposed framework outperforms the contemporary self-supervised learning framework on the writer identification benchmark and also outperforms several supervised methods as well. To the best of our knowledge, this work is the first of its kind to apply self-supervised learning for learning representations for writer verification tasks.

Earthquakes universally pose threats and risks to all inhabitants of our planet. The development of seismic sensors and earthquake alert systems stands as a crucial endeavor for society. The inherent nonstationary nature of earthquake signals presents a challenge in promptly identifying and classifying events using conventional methods such as peak ground displacement and velocity in real-time scenarios. In this research paper, we introduce an extraction method focused on first and rapidly (fast) arriving P- wave signals, providing parameters for earthquake event characteristics, specifically Pd and τc. These identified features serve as benchmarks for evaluating the efficiency of various popular machine learning (ML) classifiers in triggering alarms for an earthquake early warning system (EEW). Our paper introduces ensemble architectures for classification purposes that outperform existing classifiers and even surpass some established ensemble classifiers. A significant aspect of this study lies in the successful implementation of our proposed machine learning model architecture on the PYNQ Z2 System on Chip (SoC) FPGA development board, enabling real-world deployment. Leveraging the computational capacity of the PYNQ Z2 (SoC) FPGA board allows us to analyze signals in terms of data points, facilitating the creation of a real-time alarm triggering system.

Seismic sensing and generation of earthquake alarm is an important application for society at large. In this paper, we propose the strategy of extracting earthquake event features parameters τc and Pd from fast-arriving P-wave signals. The said features are used to explore the performances of some of the popular machine learning (ML) based classifiers to compare their performances in triggering an alarm for the Earthquake Early Warning (EEW) system. We explored four different ML classifiers namely Support Vector Machine (SVM), Naive Bayes, K-Nearest Neighbors (KNN), and Logistic Regression so that the best can be applied for the EEW alarm generation. We have used publicly available data from the PESMOS platform of IIT-Roorkee in this work.

The rapid technological advancement in autonomous vehicles due to the development in the field of artificial intelligence as well as rapid improvement in computing hardware has led to immense improvement in self driving technology. The hardware that is being implemented alongside a human driver is already capable of self driving on highways. With the passage of time and the collection of more data by these systems Full Self Driving will soon be a reality. This paper explores an approach to modelling FSD and addresses the industry leading FSD version available. The disruption that will be caused by AVs will be very large. ARK Invest's research shows that the global autonomous MaaS market will exceed $10 trillion in gross revenue by the early 2030s. In the United States, ARK expects the MaaS market to reach over $700 billion in sales by 2030, or roughly 30 times the size of the taxi industry today. The global trucking industry is expected to exceed $6 trillion and AVs will dominate that market. So, the impact of AVs must be foreseen and this paper gives a model for designing an AV network such that it can be implemented in vast scales.

Improvement in detection and evaluation of brain abnormality and tissues detection is an important task in brain image analysis. Diagnosis quality of brain MR of brain images hampered due to the presence of artifacts. Small abnormalities detection hampered due to the presence of skull region of the brain. Sometimes artifacts and skull have been treated as an abnormality in the automated system, and it hampers the intelligence system. Thus a computerized method requires pre-processed image as artifacts and skull removal. Pre-processing makes the image segmentation more accurate. In this paper, a pre-processing method for improvement of brain abnormality detection and diagnosis has been presented.

This abstract presents a pioneering dataset tailored explicitly for Earthquake Early Warning (EEW) systems. This dataset comprises a diverse collection of seismic data sourced from a network of sensors strategically positioned across seismically active regions. The dataset is designed to address existing research gaps by providing a standardized and openly accessible resource for benchmarking, testing, and advancing EEW technologies. This novel EEW dataset comprises multitudinous categories of earthquake early warning data, including average period (τc), peak displacement (Pd) and magnitude and also different earthquake centers information. The existing dataset has limited diversity and size. In this work, we propose a novel dataset for EEW systems which contains 1050 samples regarding average period (τc), peak displacement (Pd) and magnitude and also different earthquake centers information.