Hi! I'm Atharva Chandak
My research interests include
I am a final year Computer Science undergrad student at Birla Institute of Technology and Science, Pilani (BITS Pilani), India.
I've by worked on Deep Learning based Computer Vision and Multi-modal Learning on wide range of research and industrial projects including Action Recognition, Generalized Continual Zero-Shot Learning, Multimodal Advertisement Understanding, etc.
Currently I'm an Intern at Mila / Robotics Group @ University of Montreal working on the Continual Domain-incremental Object Detection.
I like building tools, tech and other hacky things so feel free to connect to discuss any interesting ideas or opportunities!
Experience
- Working on continual object detection for autonomous driving.
- Worked on long-range detection and tracking of aerial vehicles to autonomously avoid collisions.
- Analysed a new dataset for allowing building better models robust to out of distribution climatic conditions.
- Designed optimized models for real-time usage on the drone capturing high resolution feed from multiple cameras.
- Conceptualized and created an end-to-end AI enabled speech chatbot for answering user home lending queries.
- Integrated speech-to-text and text-to-speech APIs for allowing speech interaction with the bot.
- Implemented NLP pipeline for intent and entity extraction and achieved an F1 scores of 93.4% and 83.5% respectively.
- Authored the action server to process the user request demands and response curation for replying back to user.
- Contributed to the Pytorch Connectomics package adding cellpose model for neuron instance segmentation.
- Explored semi-supervised methods to improve upon the performance of 3D segmentation.
- Designed an end-to-end pipeline using long range affinity learning and transformers for improving model accuracy.
- Investigating various Vision-Transformer networks for fine-grain human action recognition.
- Reviewed the existing works on fine-grained action recognition and summarised the major research gaps.
- Exploring better temporal modelling techniques for improving performance on RGB frames directly instead of using optical flow.
- Worked on Generalized Continual Zero Shot Learning for various Computer Vision tasks.
- Integrated the incremental learning setup with zero shot learning for more realistic adaption of DL methods in everyday scenario.
- Extended the work to generalized, out of distribution tasks and also enable task free learning.
- Worked on texture classification of images using both traditional Computer vision with ML algorithms and Deep Learning based methods.
- Used traditional computer vision algorithms like FAST, ORB & BRISK combined with ML classifiers like SVMs, KNNs, etc.
- Extended the project to also implement simple image segmentation networks for performing texture segmentation useful for detecting cracks and faults in leather.
Projects
A GAN based model to convert low resolution images to high resolution.
A simulated drone which delivers multiple packages to their destinations, optimizing for time and quantity.
A jekyll template for easy creation of course websites.
Skills
Languages
Libraries & Frameworks
