Hiresh Gupta

Hiresh Gupta

CV Grad Student at CMU | ML@Apple | Adobe | BITS-Pilani (CS)

Universal Face Model

Advisor: Prof. Fernando De la Torre:

  • Presented solutions for 3D Face Reconstruction from a single image & used StyleGAN to perform interactive face editing.
  • Explored the use of Occupancy Networks to predict displacement maps and used ID-MRF loss to improve photo-realism.
  • Project Website: https://mscvprojects.ri.cmu.edu/2022team3/

MultiGAN Distillation

Advisor: Prof. Deepak Pathak:

  • Analyzed training strategies for effective knowledge transfer from GANs to target domains with few images.
  • Proposed solutions for data-efficient training of GANs and fusion of multiple pre-trained generators into one.
  • Project Website: https://sites.google.com/andrew.cmu.edu/multigan-distillation

Deep Product Search In the Wild

Advisor: Mr. Balaji Krishnamurthy

  • Worked on a deep learning based Visual Product Search which segments different products in a wild image and performs search in a large catalogue.
  • Proposed a novel grid based training of Siamese networks, allowing it to observe multiple positive and negative image instances simultaneously. The research was awarded the Best Paper Award at CVPR 2019 Workshop on Fashion and Subjective Search.

Automated Image Captioning using Multimodal Recurrent Nets

Advisor: Prof. Poonam Goyal

  • Designed an end to end model for generating sentence long descriptions of an input scene, using CNNs to extract image features and RNNs to decode them into natural language.
  • Used Beam Search, Scheduled Sampling and Caption re-ranking strategies to improve the caption quality.

Water Quality Analysis using Machine Learning Techniques

Advisor: Dr. JL Raheja

  • Applied statistical analysis and machine learning techniques to monitor water quality from different water sources in India.

Image Super Resolution using deep Convolutional Neural Networks

Advisor: Prof. Surekha Bhanot

  • Implemented a deep CNN from scratch to generate a higher quality/ detailed version of the low quality input image.
  • The model outperformed the traditional heuristic based approaches on Timofte benchmark. The research was also presented in a paper at APOGEE, BITS Pilani.