Google Research, Mountain View, CA
Core business contributions:
-> Large-scale data generation pipeline for reposed people.
-> Generating consistent virtual avatars of the person with the ability to repose, change garments, lighting, and background.
-> Setup JAX Flux for distributed training on TPUs.
Sandeep Mishra*, Oindrila Saha* , Alan C. Bovik
Oral @ AIGENS and Oral @ SP4V at ICCV, 2025
-> Decoupling motion, object shape, and object appearance in a single pipeline.
-> We achieve better object swapping while reducing structure leakage.
-> Our method simultaneously modifies objects and backgrounds, independently.
Paper
Sandeep Mishra*, Oindrila Saha* , Alan C. Bovik
In Proceedings of Neural Information Processing Systems (NeurIPS) 2024
Paper Webpage Code
Sandeep Mishra*, Avinab Saha* , Alan C. Bovik
In Proceedings of Computer Vision and Pattern Recognition (CVPR) 2023
Paper Code
Sandeep Mishra*, Oindrila Saha* , Alan C. Bovik
In AI for 3D Generation workshop at Computer Vision and Pattern Recognition (CVPR) 2024
Paper
Sandeep Mishra*, Shashank Gupta*, Ramit Pahwa, Margaret H. Pinson, Alan C. Bovik
Manuscript Under Preparation
Website Database
Sandeep Mishra*, Mukul Jha , Alan C. Bovik
Submitted at the IEEE Transations on Image Processing
Paper
Dataset
Sandeep Mishra* and Oindrila Saha*
British Machine Vision Conference (BMVC), 2020
Paper
Most of the platforms we use to share images like Facebook, Whatsapp and Instagram compress images while sharing to reduce data consumption. This leads to image quality degradation. To mitigate this we designed an AI Gallery zoom feature which uses an AI based upscaler to replace the current hardware scaler. When a user zooms into an image this feature improves overall image quality and sharpness without introducing any visible artifacts. The network has been designed such that it uses minimal power and compute requirements to achieve superior results. It is currently supported in almost 10 Samsung high-end smartphone devices. You can checkout more about this work here.