May 2022 – Sep 2022 · 4 mos · Santa Monica, California, United States · Hybrid

• Implemented monocular video-based 3D object detection which predicts the 3D bounding box of objects along with their velocities from a temporal sequence of images on the NuScenes dataset.
• Designed many inter-frame fusion techniques, including attention models, to improve the velocity predictions of detected objects by 40%, with just 15% computation overhead.
• Filed a US patent covering all the techniques.

Sep 2021 – May 2023 · 1 yr 8 mos · Boston, Massachusetts, United States · On-site

• Lab: Robust Systems Lab
• Advisor: Prof. Octavia Camps
• Implemented binarization module to perform non-zero thresholding which reduces noise and perspective distortion, increasing the accuracy of the model from 87.5% to 92.9%.
• Designed a pipeline that includes activity recognition and object detection, to estimate fish count using Pytorch.
• Integrated transformer-based encoder-decoder network into multi-view action recognition pipeline, increased the accuracy by 10%.

Dec 2019 – Sep 2021 · 1 yr 9 mos · Bangalore Urban, Karnataka, India

• Built a real-time object detector and tracker using center-point based keypoint detection technique in MXNet.
• Designed a camera-pose invariant 3D box object detector trained on rendered 3D CAD models using Blender.
• Increased the model inference speed by 20x with maintaining the same accuracy, by creating the network graph in ONNX and deploying it in TensorRT.

Jan 2019 – Nov 2019 · 10 mos · Bangalore Urban, Karnataka, India

• Implemented driver activity detection feature to spot activities such as smoking, using phone, eating, and drinking.
• Rendered face images for different head-poses by fitting a dense 3D face model and rotating the image in 3D space.

Apr 2018 – Dec 2018 · 8 mos · Bangalore Urban, Karnataka, India

• Worked on 3D facial landmarks detection, face tracking, eye and mouth tracking, and head-pose detection.
• Employed emotion detection feature to identify if the driver is angry, disgust, sad, happy, surprise, fear or neutral.

Apr 2018 – Jul 2018 · 3 mos · Gurgaon, India

• Increased processing speed by 40% by implementing batch processing for custom real-time object detector.
• Quantized 32 bit model to 16 bit model and 8bit model to reduce memory access cost and increase compute efficiency, trained and evaluated the models on COCO dataset.
• Used Computer Vision techniques to develop a new fully integrated functionality in the company’s security solution which is now integrated with the product after rigorous long-term testing at multiple customer locations.

Aug 2017 – Aug 2018 · 1 yr

• Trained a KUKA KR5 robot to perform the task of arranging objects when sequences of the same task are provided as an input.
• Employed Single Shot MultiBox Detector(SSD) to detect all objects in the frame in 2d space.
• Calculated location and orientation of the objects from 3d point cloud using Point Cloud Library(PCL).

Oct 2016 – Mar 2017 · 5 mos

• Built PCBs for sensor based object detection and for integrating Arduino MegaADK, Sabertooth, Encoders and Raspberry pi2 as a single module.
• Worked on face recognition, interaction with human and improved efficiency of battery usage by the entire system.
• Implemented Large Scale Direct(feature-less) Monocular SLAM algorithm for 3D reconstruction of the environment.
• Built a Maze solving two wheel driven mobile robot using line sensor, camera and PID based control on Raspberry pi 2.
• Received Seed Grant of 6000 USD from IITD for project development.

Jul 2016 – Apr 2018 · 1 yr 9 mos

• Developed computer vision module for Level 2 Self Driving Car.
• Automated traffic sign detection, lane detection as well as programmed it to avoid bumps using Tensorflow.
• Selected among the top 13 teams in the Mahindra Rise Driverless Challenge out of 600+ applicants in 2017.