Publications — Pujari Keerthika

A NeRF-Transformer Hybrid Framework for High-Quality 3D Scene Reconstruction and Contextual Interpretation

Pujari Keerthika, Padmalakshmi S, Moksha R, M.S Minu · 2025 1st International Conference on Smart and Intelligent Systems (SISCON 2025), IEEE · DOI: 10.1109/SISCON66686.2025 · IEEE Xplore

Published

Abstract

In this project, we would be exploring the integration of NeRFs and Transformers, creating a hybrid pipeline for 3D Scene Understanding. NeRFs is a novice approach to reconstructing 3D scenes from 2D sparse image inputs. However, there are limitations in spatial understanding and complex scene understanding. Transformers offer a global attention mechanism and feature extraction abilities, and hence leveraging them would improve the spatial representation and coherence of reconstructed scenes. Performance is evaluated on both synthetic and real-world datasets, and benchmarked against standard metrics like PSNR and SSIM. This project holds the capability to significantly impact applications in virtual reality, autonomous systems, and augmented reality by advancing the scalability and robustness of 3D scene reconstruction techniques.

Approach

The system reconstructs 3D scenes from 2D images by combining Neural Radiance Fields with Vision Transformers. NeRF generates photorealistic 3D environments from multi-view images, while a transformer layer (SegFormer) adds semantic understanding by detecting objects and how they relate to each other in space. The pipeline processes camera poses, generates rays, and runs volumetric rendering to produce RGB outputs and depth maps, with an interactive 360° viewer for exploring the reconstructed scene. Output quality is benchmarked with PSNR and SSIM. Built with Python, PyTorch, TensorFlow, SegFormer, NumPy, Matplotlib, and scikit-image.

Cross-Subject Robust EMG-Based Hand Gesture Recognition Using a Hybrid CNN-Transformer

Padmalakshmi S, S. Sridevi, Pujari Keerthika, Moksha R · 17th International Conference on Recent Engineering and Technology (ICRET 2026)

Accepted — publication pending

Abstract

Surface electromyography (sEMG)-based hand gesture recognition has garnered significant attention for wearable HCI, assistive communication systems, and prosthetic control. Reliable cross-subject recognition is still challenging, though, because individual users have variable muscle physiology, electrode location, and signal amplitude. This paper proposes a hybrid deep learning system that combines transformer-based temporal modeling with multi-scale convolutional feature extraction to generate discriminative representations from multi-channel EMG data. The method employs contrastive learning, that is supervised, and focus loss to improve the robustness of representation and classification performance when processing the EMG data from the MYO armband. Experimental evaluation on a collection of recordings from 36 subjects shows strong cross-subject performance, with 92.3% macro F1-score and 92.4% classification accuracy. The findings show that the suggested architecture enhances generalisation across unseen subjects and successfully captures intricate muscle activation patterns.

Approach

Signals from a MYO armband's 8-channel setup are preprocessed, windowed, and fed into a multi-scale CNN encoder with parallel branches using different kernel sizes to capture short, medium, and long-duration muscle activation patterns. An SE attention block re-weights channels so the network focuses on the EMG channels that matter for each gesture, and a transformer encoder handles long-range temporal dependencies via multi-head self-attention. Training jointly optimizes a classification head (focal loss) and a supervised contrastive head, which boosts accuracy and helps the model generalize to users it hasn't seen before.