Lakshmi Narayan Jayapal

Browser-based Background Blur Implementation

Trained and optimized an on-device person segmentation model for real-time background blur effects in web browsers. The model powers the change background feature in Zoho Cliq and Meeting.

• Architected a lightweight decoder and multi-scale feature extraction system that improved segmentation accuracy while reducing inference time and computational overhead.
• Performed QAT training to quantize the float32 model to int8 for faster CPU inference.
• Implemented an iterative semi-supervised learning for training a powerful teacher model for person segmentation using limited human-annotated data and pseudo labeled data.

Face Landmark detection

Trained a multi-task deep learning model for face detection, tracking, head-pose estimation and landmark regression. Contributions:

• Use pose-based data balancing strategy to mitigate the difficulty of localizing landmarks on faces with extreme pose variations.
• Implemented 1-euro filter to reduce jittering of face landmarks.
• Implemented loss-based and gradient-based methods for adaptive loss balancing.

Semantic entity recognition

Finetuned the LayoutLM model to extract first name, last name, policy group number, DOB,

and other entities from health insurance cards using only text and layout information.

Self-supervised backbone training

Trained a vision transformer network using unlabeled image data. Partnered with the Zoho Creator team to deploy the model as a fixed feature extractor for few-shot image classification.

• Build a data processing pipeline with image deduplication and Faiss-based image retrieval to curate a dataset of 1.4 million images.
• Distilled a low compute Vit-Small model using knowledge distillation technique.
• Implemented a logits based post-hoc scoring function to detect OOD samples.

Few-shot counting

Trained a deep-learning model to count object instances (arbitrary categories) in images.

• Programmatically curated candidate images from the Open-images dataset and used the mosaic augmentation to synthesize training images to resolve challenges of long tailed distribution.