Mrinmoy Kumar Das

Developed a Unity 3D VR environment integrating synchronized eye-tracking, head pose, and movement trajectory logging for precise behavioral data collection.

Built a Streamlit dashboard for real-time experiment control and data capture, enhancing research efficiency and analysis capabilities.

Evaluated InternVL2 for language-guided navigation, comparing RGB-only and RGB+semantic+depth prompting pipelines to analyze contextual effects on reasoning and instruction following.

Optimized a multi-loss objective by combining SegLoss, EdgeAttention, and InverseFormLoss for MobileNet V2Plus, achieving a 77.68% mIoU in assistive-navigation scene segmentation.

Built a multi-camera-1D LiDAR fusion pipeline using calibrated extrinsics to align range measurements with visual features, generating BEV representations for localization-aware obstacle reasoning and navigation.

Trained Swin Transformer segmentation models with MMDetection on a 14-node H100 cluster, scaling infrastructure parsing across 140,000+ synthetic Unity images with masks, boxes, and keypoints.

Improved real-world wire segmentation accuracy from 67% to 83% via sim-to-real transfer, utilizing geometric line fitting to merge patch-level predictions and suppress false positives.

Built a comprehensive training and deployment workflow for named entity recognition and knowledge graph generation using spaCy transformers.

Standardized customer-provided data for model training and reduced computation cost by 35%, optimizing resource utilization and project delivery.

Built and evaluated CNN baselines for brain cancer classification using VGG-16, VGG-19, and ResNet architectures.

Achieved 84% accuracy with elastic deformation augmentation, significantly improving model generalization for medical image analysis.