Rittwika Kansabanik

Ph.D

Statistics

Florida State University

PH.D. STUDENT

Research Experience [Novel Feature Selection Technique for Object Detection Problem] - [PI: Prof. Adrian Barbu] [Developed a feature selection algorithm which filters out all the irrelevant features to reduce training data requirements and to achieve better generalization for object detection problem.] • Several popular image data like ImageNey-1k, Cifar100, ilab2M-Light have been used, which contain millions of images of different objects. • ResNet 50 model has been used to extract features from image datasets and used for further evaluation. • An efficient algorithm has been developed to rank the features according to their significance based on the inverse sample covariance matrix and ensembled with Probabilistic Principal Component Analysis. • Achieved 84% accuracy in identifying the right objects. [No Reference Video Quality Prediction Using CNN & LSTM] [PI: Prof. Adrian Barbu] [Developed a model to assess the quality of videos without using any reference information. The method is based on a combination of pretrained CNN and a long short-term memory (LSTM) network.] • KonVID-1k (1200 videos) have been used which have authentic distortions and sampled from Yahoo Flickr Creative Commons 100 million. • The pretrained CNN model has been used to extract spatial -temporal information and the LSTM for capturing the temporal dependencies. Transfer learning has been carried out while training. • 82% accuracy has been achieved in prediction of video quality.