Projects

2024–2025

AI Rahman

Music Information retrieval is one of the polarizing areas in the field of AI today. A number of sequence to sequence based models have been explored for single track music generation, but bringing in multiple instruments together for the generation of multi-track music composition is yet a challenge. It requires a mechanism that maintains the coherence while generating music. By means of this project, we aim to explore models that are able to coherently fuse two different genres and create pleasant music.

Night Vision

Low light conditions often degrade the quality of images and videos, making it challenging for various applications such as CCTV surveillance and satellite imaging. This project aims to address the problem by enhancing low light videos using deep learning techniques. Specifically, we will be working with CNNs, GANs, diffusion models, object detection algorithms (eg. YOLO), temporal convolutional networks for improving the visibility and quality of videos captured in low light environments.

Suncast

High-energy solar flares and coronal mass ejections (CMEs) have the potential to disrupt Earth’s satellite infrastructure significantly. Predicting these events in advance could enable the activation of defense mechanisms to mitigate their effects. Since these phenomena often originate from regions of the solar surface associated with sunspots, predicting sunspots is also crucial. By leveraging Graph Neural Networks (GNNs) to predict sunspots, we aim to achieve enhanced prediction accuracy in this domain.

2023–2024

Text 2 Scene

The project developed a novel image editing tool that allows the user to make changes to a sample image using English text. We used masking and impainting, followed by harmonization to achieve this in a photo-realistic fashion.

Spike Drive

The project aimed to utilise Spiking Neural Networks (SNNs) to identify moving objects. We successfully implemented a Convolution SNN on the DVS Gesture Dataset achieving an accuracy of 78%.

RL Games

The project explored Reinforcement Learning Algorithms and Deep Reinforcement learning techniques such as PPO, SAC etc, in complex environments ranging from MineRL to Mario.

Deepfake Detection

The project utilises the SeqFakeFormer architecture to trace and detect deepfakes. The pipeline involves CNNs to capture feature manipulations and cross attention models to capture sequential manipulations.

OptiWing

The Project is based on AI applications in the field of Aerospace Engineering and Airfoil design. We tested out different model pipelines to generate robust and smooth airfoils given the corresponding 'Cp' vs 'X' plots.

AI Choreography

The project aimed to generate realisable dance moves for music inputs provided. We implemented a Spatial Temporal GCN and used cross-attention models to establish the relationship between audio and motion features.