Deepak Kumar
Email:
mir.em.@kkcaa3plaudrgome unscramble
|
I am a Software Engineer at Google. Prior to this I worked for Microsoft on Edge browser logging for Bing feature enhancement, at Search Technology Center India, Hyderabad. Before this I worked on flight booking experience on Bing. The experience is shipped in 200 different regions and languages. I had also interned at Microsoft in Summer 2015 working on leverageing Openstreetmap for Bing Maps after which I was offered a full time role at Microsoft. Before all this I graduated from IIT Kanpur in 2016 with Bachelors in Computer Science and Engineering. I did most of the projects under Prof. Vinay Namboodiri on computer vision. During my time here I also volunteered for NGO BloodConnect with a mission to tackle the blood shortage in Kanpur. |
|
|
|
|
report |
poster |
description |
code
Finetuned ImageNet pretrained model to adapt to pedestrian detection by changing the final CNN innerproduct layer. Used this caffe model to generate feature vector corresponding to each bounding box. Trained a LinearSVM for final classification. Obtianed overall precision of 88.2% on INRIA person dataset. |
|
report |
description
Implemented a real-time background-foreground classifier for the surveillance videos of IIT Kanpur. Used GoogLeNet pretrained model to generate feature vector for each frame and trained it on LinearSVM. Obtained an overall precision of 85% with a processing rate of 220 fps on GPU enabled machine. |
|
report |
description |
code
Created a scalable vocabulary tree for object recognition on UKY dataset. Used hierarchical k-Means clustering on SIFT features to create the tree and implemented idf based scoring. Evaluated the created vocabulary tree with different branch factors and different depths. |
|
description |
poster |
paper
Exploited multiple sources of information for object recognition tasks when additional modalities that are not present in the labelled training set are available for inference. Predicted the missing data for the labelled examples and exploited all modalities using multiple kernel learning. Performed hallucination to infer the missing modalities in the labelled dataset. |
|
video |
description
Implemented ‘Air Hockey’ game between two bots (One completely automated, other manually controlled). Studied the accuracy in positioning of the automated bot on approach of ball when hit by the manual bot. Optimized the positioning of the bot by improving algorithm for ball detection (using overhead camera) and dynamic calibration of the rpm of motors. |
|
report |
description |
code
Developed a Biometric Fingerprint based attendance system with storage of 200 fingerprints at a time. The device could be circulated in a class and the attendance automatically marked in the professor’s PC. Implemented python based user-friendly GUI and MySQL based queries for easy extraction of records for each course and individual student. |
|
description |
code
Implemented an end to end compiler for Perl in python which generates machine code for x86 architecture. Supported data types such as Integers, Strings along with operators, statements, loops and functions (recursive). |
|
description |
code
Implemented a meta circular interpreter for a declarative sequential model for Oz. Provided support for unification, record operations and pattern matching, over an abstract implementation of the single assignment store and semantic stack using easy-to-parse abstract syntax. |
|
site |
description |
code
Scraped institute’s course data of past 5 years and made it available to the students during pre-registration time. Added the list of students who opted for a particular course in a particular year so that one could get to know who to contact for a particular course’s content and structure. Received ~5000 hits in one week. |
|
|
