PCD-CNN Disentangling 3D Pose in Dendritic CNN for Unconstrained 2D Face Alignment

  • Condition on 3D pose to disentangle pose information from the image
  • A dendritic CNN for learning relationships between keypoints has been proposed
  • Mask-Softmax and Offline hard mining to achieve state of the art result on AFLW, AFW, COFW and IBUG datasets

Keypoint estimation of unconstrained face images -KEPLER

  • Developed an iterative algorithm for key-point detection on unconstrained face images.
  • Used novel deep convolution neural network architectures for structured regression problem
  • Achieved state of the art result reducing the error by more than 50%, accept as oral in Face and Gestures’2017

Head pose estimation of occluded face images by regularized regression

  • Course project for Convex Optimization, University of Maryland
  • Developed a regularized regression framework for head pose estimation of occluded face images.
  • Used nuclear norm regularized approach, to remove the occlusion, preserving the facial details

Face Alignment by Local Deep Descriptor Regression

  • As a part of Janus team, developed face alignment algorithm for the pre-processing of the dataset.
  • Using deep features extracted from the image, applied regression techniques to achieve unconstrained keypoint detections

People Detection and tracking using RGBD data (IIT Kharagpur)

  • Revamped the people detection and tracking algorithm on the RGB-Depth datasets
  • Performed mathematical modelling of human head to create templates for accurate people identification

Summer Internship (Johns Hopkins University)

  • Developed software to be used in hospitals for annotating the data for pre-processing and training in later stages
  • Implemented human detection algorithm using Point Cloud Library on the ICU dataset collected at Johns Hopkins Hospital
  • Validated the SMIJ feature descriptor for action classification on the Microsoft 3D Activity dataset

Removal of Rain Streaks from videos using spatio-temporal characteristics (IIT Kharagpur)

  • Applied Bayesian Inference, and neural networks using Spatial Temporal features to classify the pixels as rain
  • Achieved accuracy up to 90% on the datasets of video clips available open source

Summer Internship (KTH Royal Institute of Technology, Sweden)

  • Proposed Noisy kernel for Gaussian Process Regression to improvise mapping of human speech segments
  • Modified the GPLVM software with the noisy kernel for smooth regression to test it on KTH Text to Speech Database