Sai Abhishek Siddhartha Namburu
Education
- Master of Science in Mechanical Engineering, Carnegie Mellon University, Concentration: Robotics and Control, 2021-2023 [current GPA: 4.0/4.0]
- B.E. (Hons.) Mechanical Engineering, Birla Institute of Technology and Science (BITS) Pilani - India, 2015-2020 [GPA: 8.74/10]
Work experience
- May 2022 - August 2022: Computer Vision Intern
- Location: SRI International, Menlo Park, CA
- Description:
- Developed an Attention-based Visual Dossier matching module to recognize whole-body ID with an accuracy of 91% based on 3D mesh and view angles generated using VIBE from pixel enhanced videos.
- Generated pixel enhanced videos (2x better SNR) from U-Net based network by removing pixel shift and blur in videos.
- Trained VIBE with an additional silhouette-based loss to generate shape aware 3D meshes of humans from videos.
- August 2020 - August 2021: Machine Learning Engineer
- Location: DeepEdge, Hyderabad - India
- Description
- Collaborated to develop a face recognition solution(RetinaFace + ArcFace + SORT) for body cameras worn by security officials. Integrated the pipeline by developing an application in K7 Camera(Ambarella CV25 chip) to inference on real time data. [code] [code]
- Modified and trained a Mask-RCNN to detect custom objects, draw masks and identify keypoints of interest in pytorch. Improved inference speed by 3x by converting the model into TensorRT using TensorRT wrappers.
- Accelerated Iterative Point Cloud algorithm (based on open3d) by 2.5x using cuda kernels and thrust library.
- Analysed and calculated GAIT Characteristic parameters like stepcount, stance, swing from a pre-trained openpose pose detection model for an autonomous walker.
- July 2018 - December 2018: Fintech Intern
- Location: Nomura Services India Private Limited, Mumbai - India
- Description
- Assisted Nomura’s FinTech division: Responsible for conceptualization and execution of Nomura technology development program - Implemented AI/ML, NLP technologies to automate use-cases across various divisions.
Research/Academic experience
- September 2021 - Present: Graduate Research Assistant
- Location: Human Sensing Lab, Carnegie Mellon University
- Description:
- Modified current deep mind based I3D action recognition model(with non-local blocks) to localize activity and achieved an accuracy of 72% (in real domain) training on combination of synthetic and real robot cognition data.
- Implemented Gradient Reversal Layer for I3D improving accuracy in real domain by 2% without using labels in real domain while training.
- Code
- Supervisor: Professor Fernando De La Torre and Prof. Jessica K. Hodgins
- August 2020 - August 2021: Project Assistant
- Location: Indian Institute of Science, Bangalore - India
- Description
- Devised optmisation algorithms in python for efficient heat transfer.
- Supervisor: Professor Anathasuresh, Prof. Pramod Kumar
Projects
- Inter-Vehicular Depth and Velocity Estimation using a Monocular Camera (CMU): February 2022 – May 2022
- Description:
- Modified Monoflex to regress for 3D key points from single image to predict depth of vehicles. Reduced dependence on data by implicitly learning for uncoupled representations (3D orientations) of a car maintaining similar AP.
- Integrated kalman filter based AB3MOT with the Monoflex (using its depth predictions) achieving 3% less error in velocity prediction compared to using PointRCNN backbone.
- Attained comparable accuracy by regressing for velocity and depth directly from a Multi-scale Vision Transformer.
- Tech Stack: Python, PyTorch
- Code and Report
- MyTorch Framework and Deep Learning (CMU): February 2022 – May 2022
- Description:
- Developed a deep learning framework with functional modules like Conv1D, Conv2D, LSTM, GRU, Attention and iterative optimization algorithms including Adam, SGD and RMSprop.
- Tech Stack: Python
- Code and Report
- Visual Learning and Computer Vision Assignments (CMU): August 2021 – May 2022
- Description:
- Implemented a weakly supervised object localization network and a deep detection network with AlexNet backbone on Pascal dataset and obtained comparable accuracy with supervised object detection methods.
- Trained a GAN on CUB 2011 dataset to generate realistic looking birds with standard GAN, LSGAN and WGAN-GP losses and compared the FID scores. Tested disentanglement of latent space by interpolating it.
- Implemented visual object tracking with Lucas-Kanade forward additive image alignment algorithm and Matthew-Bakers Inverse Compositional alignment algorithm.
- Performed 3-D Reconstruction of a temple estimating Fundamental matrix using with 8/7-point algorithm and RANSAC.
- Worked on stitching panoramas utilizing homography and RANSAC.
- Implemented MNIST character detection recognition using neural nets and acheived an accuracy greater than 90%.
- Used Spatial Pyramid Matching of image word maps to recognise the location of an image.
- Tech Stack: Python, PyTorch
- Code and Report
- Emotion Recognition in Multi-Party Conversations using Multi-Modal Approach (CMU): October 2021 – November 2021
- Description:
- Achieved an accuracy of 60% in predicting sentiment and emotion by combining text and audio modules of multi-party conversations in friends using shallow machine learning models.
- Deployed various feature engineering methods to improve the accuracy by 5%. Compared the results with t-CNN, d-RNN and bc-LSTM and reasoned the accuracies.
- Tech Stack: Python, PyTorch, Tensorflow
- Code and Report
- Robot Planner Visualization (CMU): October 2021 – November 2021
- Description:
- Implemented RRT, RRT, RRT Connect, A, Weighted A* and Dijikstra in 2D using object oriented paradigm.
- Visualized the search of these planner using openGL in 2D. Developed a GUI to make the product market ready.
- Tech Stack: C++, openGL
- Code and Report
- Home Service Robot: June 2021 – July 2021
- Description:
- Deployed a turtlebot mobile robot in a designed custom environment in gazebo to pick-up and drop-off objects in the environment.
- Mapped the environment using gmapping package, localized it and used ROS navigation stack (Dijkstra’s algorithm) to plan a safe path to pick-up and drop-off objects (designed with rviz marker).
- Tech Stack: C++, ROS
- Fashion Retail forecasting using ENN: February 2019 – April 2019
- Description:
- Modelled custom evolutionary neural network using genetic algorithm to forecast demand for two different apparels.
- Employed a pre-search algorithm along with BIC metric to effectively increase computation speed.
- Compared with a Non-linear Auto Regressive exogenous input network to pick a better forecasting method.
- Tech Stack: Python, Keras, Scikit-learn, Matlab DL tool
- Code
- Classification of YouTube comments to detect advertisements: November 2017 - December 2017
- Description:
- Implemented an end-end approach with Support Vector Machine and a Naive Bayes classifier to classify YouTube comments and detect advertisements based on words used.
- Used RoC-AuC metric to tune hyperparameters in the model.
- Tech Stack: Python
- Code
- Algorithm to reduce epicyclic geartrains by Isomorphism test: February 2018 – April 2018
- Description:
- Devised an algorithm to reduce the number of epicyclic gear trains after checking structural and rotational isomorphism.
- Algorithm first generates all possible Epi cyclic gear trains(EGT) and associates each them to a matrix using hammingmethod, then checks for translational and rotational isomorphism retaining non-isomorphic gear trains.
- Tech Stack: Matlab
Skills
- Programming Languages: Python, C/C++ (Cuda kernels and thrust library), R
- Deep-Learning Frameworks: PyTorch, TensorRT, Keras, Tensorflow(1 and 2)
- Tools: Docker, Scikit learn, OpenCV, SQL, MS Office, LaTeX, GitHub
- Application Software: ROS, MATLAB, Gazebo
Publications
Research & Teaching
Certificates & Achievements
- Robotics Software Engineer : Udacity [In Progress]
- Deep Learning Specialization: Deeplearning.ai, Coursera [December 2018]
- Machine Learning: Coursera [December 2018]
- C, C++ programming: NIIT [May 2019]
- Excel to My SQL Analytics techniques: Coursera [Jun 2018]
- Bloomberg Market Concepts: Bloomberg [Aug 2017]
- R programming training: Innodatatics USA [Aug 2017]
Activities and Interests
- Secretary at Brindavanam Telugu Samithi at BITS-Hyderabad
- Member of VFX Club at BITS-Hyderabad
- Member of Department of Visual Effects at BITS-Hyderabad
- Teaching Volunteer for nearby schools in under-developed areas as a part of Nirmann, BITS Hyderabad
