Optical Flow Estimation using Graph Convolutional Networks
Summer Research Fellow at the Computer Vision Imaging and Graphics (CVIG) Lab, Indian Institute of Technology (IIT), Gandhinagar, India
Guide: Prof. Shanmuganathan Raman, Indian Institute of Technology, Gandhinagar, India
Implemented an encoder-decoder based Graph Convolutional Network (GCN) framework for the task of Optical Flow estimation. Performed experiments using various GCN models like GCNII and DeepGCNs(ResGCN and DenseGCN) on the MPI Sintel and KITTI datasets. Validated the effectiveness of the GCN learned representation model on frame order prediction by taking temporally shuffled frames (i.e., in non-chronological order) as inputs.
CoviBioBERT : A pre-trained Named Entity Recognition Model in Biomedical Domain
Research Internship at the AI and NLP Lab, Indian Institute of Technology (IIT), Roorkee, India
Guide: Prof. Raksha Sharma, Indian Institute of Technology, Roorkee, India
Worked on the implementation the CoviBioBERT model, where the COVID-19 open research dataset was used for pretraining using weights of the BERT model. Trained the model for 100K steps for a maximum sequence length of 128 and further trained it for additional 25K steps for a maximum sequence length of 256. Concluded that the CoviBioBERT model can be very useful for NER specific task in COVID-19 domain in future as the accuracy obtained from this model was close to that of the BioBERT model.
Few-Shot Learning for Visual Question Answering
Currently a research Intern at the the Video Analytics Lab(VAL), Indian Institute of Science (IISc), Bangalore, India
Guide: Prof. Venkatesh Babu, Indian Institute of Science (IISc), Bangalore, India
Analysed various few-shot learning algorithms like Matching networks, Model-Agnostic Meta-learning and Prototypical networks for the task of Visual Question Answering (VQA). Currently working on applying Open Long-Tailed Recognition (OLTR) algorithm for few-shot learning VQA.