Worked with AirSim simulator to generate synthetic data to test object detection models on new environments.

Used a physics-based particle simulator in order to augment clear weather conditions to rain and fog of varying intensities.

Achieved around 55% improvement in mAP score after training YOLOv5 on augmented weather data.

Implemented state-of-the-art monocular depth estimation and instance segmentation algorithms using PyTorch.

Incorporated scale and shift factors to arrive at quantitative estimates of depth from the estimated disparity map.

Used Delaunay Triangulation and projection-based volume estimation to develop a volume estimation application.

The % error averages around 20% for regular objects (such as ball) and 40% for irregular objects (such as heaps).

Devised a theoretically rigorous criterion based on hypothesis testing to estimate order of the MIMO model.

Developed a novel iterative algorithm based on Principal Component Analysis in MATLAB that estimates measurement error variances and state space model parameters in the presence of process and measurement noises.

Simulation experiments indicate that the proposed algorithm provides consistent estimates of all parameters.

Assisted in developing a Residual Generative Adversarial Network consisting of Bidirectional Convolutional LSTMs.

Ideated the incorporation of an attention mechanism to dynamically weight high and low-frequency frames for fusion.

Developed model achieved state-of-the-art results for both Video Super-Resolution and Deblurring tasks

Incorporated divide-and-conquer algorithm to train a Wide and Deep Residual Neural Network using TensorFlow.

Designed a learnable Pixel Calibration layer that computes relative importance of pixels to enhance overall efficiency.

Achieved state-of-the-art performance through the proposed divide-and-conquer approach for 2x, 3x and 4x scaling.