Manoj Srinivasan

I am an MS Computer Engineering Student at New York University, having done my undergraduate studies at IIT Madras. I enjoy working on deep learning research specifically in the area of computer vision.

Email  /  CV  /  Twitter  /  Github

Aimed at generating counterfactuals conditioned on the initial image, and a text prompt to control its several attributes while minimally altering the image. We propose to perform image inpainting on regions corresponding to high confidence towards the target class using Stable Diffusion and CLIP models.

Implemented an implicit 3D reconstruction neural network that constructs a surface mesh using its point cloud data. The network is trained to find the occupancy metric of different point coordinates in the point cloud, in turn finding the points present on the surface during inference.

Involves performing semantic segmentation on pictures of steel surfaces to classify and localize the surface defects using the UNET architecture. Got an average of 95 - 97% IOU score.

Simplified SOTA generative networks namely StyleGanv1,2 to lower network capacity and introduced dense connections between multiple resolutions. Improved perceptual image quality scores by 10% even with as little as 100-1000 training images.

Website template borrowed from Jon Barron