Statement of Purpose

I am applying to the M.S. CS (Thesis) program to further explore computer vision for autonomous systems and to
develop an advanced foundation in AI through graduate research and coursework in preparation for a career in applied
research. My research interests in multimodal perception, scene understanding and adapting to challenging scenarios for
AVs have been shaped by my research experience in domain adaptation and multi-task learning in computer vision. The
advancements in domain adaptation and self-supervised learning present exciting avenues for improving deep vision and
I seek to explore such applications at UCSD for my master’s thesis. My ultimate goal is to make valuable research
contributions in Computer Vision, and the M.S. CS program is the ideal next step because the master’s thesis and
the excellent vision research at UCSD present the perfect environment for deep exposure to AI.

My interest in Computer Vision stems from a project with Inkers AI. The exposure to exciting applications of
video intelligence and edge-deployments sparked my curiosity, and to build a strong foundation in AI, I picked electives
such as ‘Digital Image Processing’ and ‘Artificial Intelligence’ for my Bachelor’s in Computer Science and have also
completed several online specialization courses. In addition to pursuing research internships, I have participated and won
several national-level competitions such as the ‘NASA SpaceApps Challenge’ and the ‘AICTE Vishwakarma Award’ for
impactful applications of vision and AI. These experiences have enabled me to understand the potential for positive
impact in Computer Vision and has oriented me towards pursuing a career in applied research.

I decided to get research experience in Computer Vision and pursued full-time research at the IISc Video
Analytics Lab. I worked under the guidance of Prof. Venkatesh Babu on ‘Multi-domain Image Translation for
Extreme Weather Degradation’ to develop a novel image translation model to effectively disentangle domain-invariant
content and domain-specific style from extreme-weather driving datasets to perform multi-domain translation with a
multiscale patch-wise contrastive loss for Indian driving datasets (domain-shift). Translating clear-weather scenes to
extreme weather scenes preserves the segmentation labels, and training on such data helps generalize to real-world
extreme weather. The results improved upon baselines and are to be reviewed at a top tier vision conference.
Working with the Ph.D. students at IISc has been immensely beneficial to my research methodology and has given me a
deeper understanding of solving real-world objectives. I have also worked on the implications of challenging weather
degradations in urban mobility and the work has been presented at AAAI 2021, UM Workshop.

For GSoC 2020, I worked with Google Brain’s TensorFlow team on ‘deep generative models to translate
night-time scenes to clear-weather daylight scenes’ under the guidance of Dr. Hongkun Yu and Jaeyoun Kim. The
purpose of the research is to improve segmentation performance and reuse labels for night-time domain adaptation. I
implemented several research experiments including energy minimization with graph cuts and attention-based
translation models and obtained the best results with a forked translation architecture. In improving the model’s
cross-cycle consistency and including a content adversarial loss and a mode-seeking regularizer, I improved translation
artifacts from rainy-reflections and glare from oncoming traffic. The experience was incredibly rewarding because of
the guidance I received from the research scientists at Google Brain in effectively formulating the research objective and
improving on artifacts and corner-cases. There was a steep learning curve in implementing the extensive experiments,
and meeting these rigorous standards gives me the confidence that I can work in research roles in computer vision.

At IISc GCDS Lab, I worked on ‘candidate policy search to find the optimal data augmentation strategy to
train crowd-counting models’. The purpose is to train vision models for drone-based flood emergency mapping to aid
response planning. I worked on candidate policy exploration-exploitation with Bayesian optimization to improve
the search-strategy for stationary augmentation policies for improved efficiency across baselines. This research gave
me a deeper insight into RL for policy schedules and efficiently training crowd-counting models. Further, I worked as a
research intern on the ‘DSFS Climate Change project’ with Prof. J. Adinarayana at IIT-Bombay to improve
UAV-vision research at scale. Working with the drone-captured multispectral data has given me a broader perspective of
the applications of drone-based vision. These experiences have helped me refine my research interests in vision for AVs.
 Working with the Ph.D. students at IISc and IIT-Bombay has been immensely beneficial to my research
methodology. The year I spent on full-time research has given me a very realistic experience of the rigors involved
in doctoral research. I enjoy the responsibility and the intellectual and creative freedom in research and believe
the Master’s thesis at UCSD will enable me to explore and focus on a depth in computer vision.

My full-time research at IISc Video Analytics Lab and work with IISc GCDS Lab and Google Brain are in strong
alignment with the research at the UCSD Centre for Visual Computing in related directions in vision for AVs. I am
interested in working with Prof. Manmohan Chandraker in scene understanding and domain adaptation for
autonomous driving. Prof. Manmohan Chandraker’s work on universal semi-supervised semantic segmentation has
informed my current research on domain adaptation for extreme weather scenes to bridge the synthetic-real domain gap
when few paired real samples are available for autonomous driving. My experience in domain adaptation and raw sensor
data from my CVPR 2021 submission: ‘few-shot domain adaptation for low light raw image enhancement’ has
suitably prepared me to work with Prof. Manmohan’s research group. I am also interested in working under the guidance
of Prof. Hao Su on 3D scene understanding applications for robotic vision. Specifically, I am interested to work on dense
optical expansion for 3D optical scene flow and research methods to improve dynamic 3D perception with
monocular image pairs with Prof. Hao Su’s group. Having contributed to the TF Graphics module for 3D mesh
segmentation on the shapes3d dataset, I have a suitable background to work on 3D scene understanding tasks.
Advancements in vision for AVs will enable fully autonomous vehicles and is an exciting research direction. I want to
make valuable contributions toward AVs and seek to orient myself toward a PhD by completing the master’s thesis.
Working on research objectives at the Centre for Visual Computing will further my understanding of computer vision.

A career spans 80,000 hours, and I want to spend that time working on some of humanity’s most pressing
problems and move the needle toward a better future. It is indeed true that the days are long and the decades are short,
and the exponential technology of today - AI - is at its tipping point and has enabled us to work on challenging problems.
I believe that an advanced foundation in AI, interdisciplinary application in fields such as robotics and
autonomous vehicles and the fantastic peer group at UCSD is the perfect enabler for me to work on challenging
research objectives. My contribution to the MS CS class is my Computer Vision perspectives from research (IISc VAL,
IISc GCDSL, IIT-Bombay), from industry (Google Brain [GSoC], CSIR-CDRI), and from cutting-edge startups (Inkers
AI) towards collaborations across course projects and research. The professors at UCSD CSE are some of the top minds
in the field, and their guidance will be instrumental in helping me develop an advanced foundation in AI. In that regard,
the MS CS with Thesis program at UCSD is an excellent fit for my skill set and experience because of its amazing
coursework, opportunity for research, and a peer group who can equip me with the skills to attain my career objective.