Jul 2021 – Nov 2021 · 4 mos · India

• BlinkPe enables in-store shopping on a phone, adding value and convenience while eliminating delays.
• Please visit our website (https://www.blinkpe.com) to know more!

Feb 2021 – Jul 2021 · 5 mos · Remote

• Worked on learning parametric 3D surface models of objects from images. Joint collaboration with colleagues from Brown, Stanford, Adobe London and IIT Kharagpur.

May 2020 – Jul 2020 · 2 mos · Mumbai, Maharashtra, India

• Goal: Use unlabelled video data to improve 3D human reconstruction and 2D keypoint prediction performance
• Conducted an intensive literature review of self-supervised techniques in several domains such as images, videos and 3d reconstruction
• Contributed to the development of a novel method for using unlabelled videos for 3D human reconstruction in a self-supervised manner
• Formulated a self-supervised reconstruction pipeline by exploiting the disentangled nature of human pose and appearance
• Improvement in 2D keypoint detection performance over supervised benchmark models such as Human Mesh Recovery (HMR)

Jan 2020 – Dec 2020 · 11 mos · India

• Goal: To develop assistive technology for visually impaired people by answering natural language questions about images
• Carried out an extensive survey of shortcomings of existing VQA models and implemented state-of-the-art models like BAN, MFB, MCAN etc
• Developed a student-teacher method to improve image attentions of any existing attention-based visual question answering system
• Gained an improvement over recent benchmarking models MFB and BAN respectively, without using more amount of data
• Devised a novel angle-based gradient update method for simultaneous optimisation of multiple objective functions

Jan 2020 – Mar 2020 · 2 mos

• Goal: To classify audio sounds as categories of sneezes for early prognosis of medical conditions
• Developed an audio classification model for a hospital to identify sneeze audios and to classify the real sneezes as soft, loud etc. for early prognosis
• Investigated the performance of image-based models and different audio features such as MFCCs, Chroma features, Log Mel Spectrogram etc.
• Obtained an accuracy of 83% (sneeze/non-sneeze), 80% (real /fake) and 60% (soft/loud/stifled/dramatic) using a CNN-LSTM architecture
• Made the model data efficient using transfer learning from a pretrained VGGish model and using heavy data augmentation
• Deployed an inference pipeline of the model using Flask and Azure Virtual Machine

Apr 2018 – Jan 2020 · 1 yr 9 mos · Kharagpur Area, India

• Autonomous Ground Vehicle Research Group (AGV) is a group of enthusiastic undergraduates working towards the end goal of making India's first autonomous vehicle.
• While we're not completely there yet, we have made substantial progress in a lot of relevant areas.
• Our group also won the 2nd prize at the Intelligent Ground Vehicle Competition (IGVC) at Michigan, USA in both 2018 and 2019.
• We are also one of the top 11 team chosen for the task of converting an electric Mahindra e2o to a fully autonomous vehicle.