Vishal Pani Email & Phone Number

Lausanne, Vd, Ch

Lausanne, Vaud, Switzerland

Master thesis on integrating neural mesh reconstruction techniques into a photogrammetry pipeline.

France

Research Intern in the GraphDeco group.Investigated the feasibility of debiasing neural rendering algorithms.Researched in connecting physically-based differentiable rendering with point-based neural rendering method like 3D Gaussian Splatting.

Continuing the summer project with an aim to reproduce results of Differentiable Rendering papers like Large Steps in Inverse Rendering of Geometry (https://rgl.epfl.ch/publications/Nicolet2021Large) from real-world captures.

Lausanne, Vaud, Switzerland

Summer internship with the Realistic Graphics Lab (RGL) led by Prof. Wenzel Jakob.• Calibrated the camera on a goniophotometer (https://rgl.epfl.ch/pages/lab/pgII) in order to match real-world captures with Mitsuba renders. • Optimized for the pixel reprojection error by improving the camera model and using image-level transformations.• Resultant pipeline reduced the pixel reprojection error from an initial average of ~10 pixels to ~1 pixel.

France

Researching Multi-Object Tracking and Re-ID methods to study interactions between Trichogramma wasps.• Developed a robust two-stage online tracker to handle the erratic movements of generic small-scale objects, in our case, Trichogramma wasps. The tracker is also extendable to any scenario where objects may enter or exit the scene. • Designed a weakly supervised sampling technique for training the ReID model, due to the unavailability of ground truth (GT) tracking annotations.• By overcoming the problems posed by erratic movement and sparse features, the tracking pipeline outperforms SOTA insect and animal tracking softwares on MOT metrics, specifically, MOTA, Identity (ID) switches, and fragmentations.

• Research internship with Dr. Norbert Haala.• Developed data preparation and deep learning pipeline for Semantic Image Segmentation of City Meshes using 2D Projections of the 3D Meshes.• Data preparation pipeline uses ray casting to generate 2D projection from the 3D city meshes.• The entire data preparation pipeline is optimized using parallel processing.• Deep Learning pipeline contains a suite of semantic segmentation models like U-Net, DeeplabV3+, etc. It also contains implementations of various weighted loss functions, optimizers, and LR schedulers

Kolkata Area, India

Investigated generative models and stabilized the training procedure of LS-GANs.● Conducted a thorough literature survey on generative models, specifically auto-encoders and GANs.● Stabilized the error oscillation of discriminator of a LS-GAN by adding a damping term to its loss function.● Improved the structural-integrity of the output of a GAN by implementing a “pre-training” scheme.● Both the experiments showed qualitatively better or competitive results to base models.

Master'S, Computer Science, 5.52/6

Bachelor Of Technology (Honors), Information Technology, 9.44/10