As a Junior Research Fellow at one of India's best research institutes, my current focus involves the development of object detection and tracking solutions using UAVs for applications of traffic monitoring and behavior analysis. The ultimate goal is to create an end-to-end model optimized for deployment on edge devices, Nvidia Jetson. Along with this, I am involved in building representation learning models for label-free tasks.

In my role as a Research Intern at ModECI (Model Exchange and Convergence Initiative), a collaborative effort among multiple investigators dedicated to establishing a standardized format for exchanging computational models across various software platforms and scientific domains like neuroscience, machine learning, and artificial intelligence, my tasks included developing serialization and deserialization pipelines. Specifically, I focused on adapting PyTorch vision models to standard format, enabling seamless interchange in JSON format for enhanced model interoperability.

GitHub - https://github.com/ModECI 

Channel and spatial attention mechanisms introduced by earlier works enhance the representation abilities of deep convolutional neural networks (CNNs) but often lead to increased parameter and computation costs. While recent approaches focus solely on efficient feature context modeling for channel attention, we aim to model both channel and spatial attention comprehensively with minimal parameters and reduced computation. Leveraging the principles of relational modeling in graphs, we introduce a constant-parameter module, STEAM: Squeeze and Transform Enhanced Attention Module, which integrates channel and spatial attention to enhance the representation power of CNNs. Additionally, we introduce Output Guided Pooling (OGP), which efficiently captures spatial context to further enhance spatial attention. We extensively evaluate STEAM for large-scale image classification, object detection and instance segmentation on standard benchmark datasets. STEAM achieves a 2% increase in accuracy over the standard ResNet-50 model with only a meager increase in GFLOPs.

arXiv: https://arxiv.org/abs/2412.09023 

Object detection and tracking tasks have been pursued by many researchers for a very long time, from traditional computer vision techniques to advanced deep learning architectures. Various object detection and tracking models have been developed for unmanned aerial vehicle (UAV) applications. However, to our knowledge, no study has yet provided a comparative analysis of existing detection and tracking models and their combinations. In this study, our focus is on implementing various object detection and tracking models for edge device deployment on UAVs. Building detectors for unmanned aerial vehicle platforms remains a challenging task and no study compares these models with UAV and edge feasibility in perspective. In this paper, we combined various object detection algorithms with different multi-class multi-object trackers to track multiple targets from the video feed and test performance on edge-device. With this comparison, we achieved a comprehensive analysis of current state-of-the-art tracking and detection algorithms that best suit the UAV use cases with different applications.

International Conference on Vehicular Electronics and Safety (ICVES) 2024 (proceedings underway)

This paper presents an efficient solution for weed classification in agriculture, focusing on optimizing inference performance while respecting agricultural constraints. It introduces a Quantized Deep Neural Network model that classifies nine weed classes using 8-bit integer quantization, reducing the model size and inference time while maintaining accuracy. The study evaluates this approach on ResNet-50 and InceptionV3 architectures, showing significant reductions in model size and inference time in real-world scenarios on processors including Desktop, Mobile, and Raspberry Pi. This work offers a promising direction for efficient AI applications in agriculture with broader potential uses.

International Conference on High Performance Computing, Data, and Analytics (HiPC) 2024

arXiv: https://arxiv.org/abs/2310.18921 

Contributed to Keras which is a prominently used open-source library for computer vision research for modular computer vision components that work natively with TensorFlow, JAX, or PyTorch.

GitHub (stars 852)  - https://github.com/keras-team/keras-cv 

Contributed to one of the most commonly used machine learning models in object detection's open-source library for easy and beginner-friendly implementation. From preparing the dataset to annotations, to model training everything is covered. 

GitHub (stars 633)- https://github.com/AntonMu/TrainYourOwnYOLO 

Contributed to the TARDIS repository used extensively by astrophysicists researchers to simulate supernovas. TARDIS packages provide several tools for physics calculation and visualization to make your supernova research easier.

GitHub (stars 182) - https://github.com/tardis-sn/tardis