NeRF: Representing Scenes as Neural Radiance Fields on a Custom Dataset Implemented the paper NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis from scratch on a custom dataset Structure from Motion – 3D Reconstruction via Bundle Adjustment Feature extraction, matching, P-n-P and iterative bundle adjustment for sparse 3D point cloud reconstruction from multi-view images Task and Motion Planning for Block Manipulation using STRIPS in Genesis Full TAMP pipeline combining PDDL task planning with OMPL-based motion planning for block manipulation in simulation enabling execution–grounding–replanning Finetuning DINOv2 with LoRA for Image Segmentation Finetuned DINOv2 encoder weights using Low-Rank Adaptation (LoRA) and a 1×1 convolution decoder, achieving 74.6% Mean IoU on Pascal VOC with only 0.69% trainable parameters. Vision-Based Robot Manipulation (VBRM) – Operational Space Control Real-time vision-based control framework for robots Autonomous Panaroma stitching using Deep Homography Net Classical panorama stitching pipeline from scratch using Harris corner detection, ANMS, SSD matching, RANSAC and compared it with the deep homography network Control and Velocity Kinematics – 4DoF OpenManipulator-X Robotic Arm Cartesian Velocity Control, Jacobian-Based IK, and PD Control Probability based boundary detection using PB-lite Probability based boundary detection algorithm considering the texture, brightness and color discontinuties in addition to pixel intensities using deep learning models like ResNet, ResNeXt and DenseNet. Einstein Vision A Tesla FSD-inspired 3D visualization pipeline for autonomous driving scenes Camera Calibration – Intrinsic and Extrinsic Parameter Estimation Zhang's checkerboard-based calibration method for recovering camera intrinsics, lens distortion, and extrinsic pose parameters Targeted Universal Adversarial Patch Attacks on Convolutional Neural Networks Targeted universal adversarial patch attacks against a ResNet-18 classifier trained on the CIFAR-10 dataset using Expectation over Transformation (EoT) Semantic Segmentation using Self-Attention weighted U-Net Self-attention enhanced U-Net architecture in PyTorch with data augmentation, improving IoU and Dice scores over baseline models
