Soham Shinde

Built an end-to-end pipeline for training and deploying segmentation and detection models for indoor navigation, packaging deployment deliverables for the Autonomous Navigation team. Migrated a complex perception system to the Qualcomm DragonWing QCS6490P chipset and NVIDIA IoT kits, improving edge inference speed by ~30% in barebones Linux environments using GStreamer. Also designed and deployed a custom 6DOF pose estimation proof-of-concept pipeline to ensure reliable target locking and rebasing during the hardware migration.

Developed and benchmarked a Vision Transformer (ViViT) for classroom activity recognition on a 927-clip EduNet subset, achieving 88% test accuracy by fine-tuning a Kinetics-400 pre-trained model with a 224x224 video preprocessing pipeline. Validated generalizability on an independent 100-video dataset (72% accuracy) and generated gradient-based saliency maps to confirm focus on key regions (e.g. raised hand, board writing). Analyzed gaze and activity patterns to derive quantitative student engagement measures.

Implemented and optimized a PyQt5-based GUI integrating SAM for image segmentation, enabling efficient AI-assisted annotation. Developed and automated a mechanism to aggregate individually segmented regions into larger sub-scenes using advanced image processing techniques, advancing the semantic understanding of complex images and reducing manual effort.

Worked under the supervision of Dr. Dhiraj Sangwan on restoring and segmenting deteriorated Rajasthani murals using deep learning models like U-Net++, DeepLabV3+, PSPNet, and FPN. Developed a synthetic damaged-image generation pipeline using GANs (StyleGAN2-ADA) and crafted diverse binary masks to simulate complex real-world mural damage. Implemented a state-of-the-art inpainting pipeline achieving an SSIM score of 0.9812 for reinstating missing sections.