Hello I’m Soumyadip, I am presently working as a Sr. Computer Vision Engineer at OpenCV University.

Previously, I worked as a Research Engineer at Infosys CAI, IIIT Delhi, focusing on Perception for the Autonomous Vehicle project ALIVE, under the guidance of Dr. Saket Anand and Dr. Sanjit Kaul. My main contributions were to the ADAS stack, HD-Map generation pipeline, Traffic Light Following (Detection, Planning & Deployment), and the development of a Virtual Testbed for AV testing.

I graduated on 2022, with a Bachelor of Technology (B. Tech) degree in the field of Electrical Engineering from IEM, Kolkata. During my graduation I worked as a research intern at IIT KGP under Dr. Debashish Chakravarty for landscape segmentation on satelite data and at Uni. of New South Wales(UNSW) under Dr. Tanmoy Dam worked on prior learning for GAN, generative clustering. I’m also a Kaggle Competition Expert, I predominantly participated in Object Detection and Segmentation competitions.

🤹 Skills: Python, C++, Carla, ROS1/2, PyTorch, NeRF Studio, COLMAP, Eigen,
                  Ceres-Solver, LiDAR, RGBD Camera etc.

📬️ Email ID: soumya997.sarkar@gmail.com

💡 Open to opportunities [research or Eng] in field of:
Computer Vision, Robotics - Perception and SLAM, 3D Vision and Graphics, Deep Learning


Projects:

Traffic Light Following

Implemented Traffic Light Following using C++ ROS package for Planning stack Behavior Tree. Trained YOLOP model for traffic light detections and Ported the model to Nvidia Orin. Integrated the model with the traffic light following stack. Conducted comprehensive testing of the P&C stack with Traffic Light Following in both the Carla and E2O platforms.

ADAS: Forward Collision Warning (FCW)

Designed and developed ADAS system, Forward Collision Warning (FCW) with Integrated Advanced Emergency Braking Systems (AEBS) maintaining the AIS standards. The project involved YOLOP object detection, tracking, and applying kinematic laws. The pipeline was tested in Carla and later deployed in Nvidia Orin.

XR-Testbed: Virtual AV testing

Worked on XR test-bed for autonomous vehicle testing. This involved RViz-based interface for efficient spawning and visualization of static and dynamic obstacles.

ADAS & DSM: Traffic Light Warning on Carla with DSM

Designed and developed ADAS system, Traffic Light Warning (TLW). The project involved YOLOP object detection, tracking, GMM Classifier, and applying kinematic laws. The following video is a showcase of the TLW and DSM on Carla. Mediapipe Facial Landmark detection was used for Drowsiness prediction and driver attention prediction.

Kaggle: TensorFlow - Great Barrier Reef [TOP 3%, Silver medal]

Problem statement was to accurately identify starfish by building an object detector trained on underwater videos of coral reefs. Trained FasterRCNN, YOLOv5. Introduced CLAHE-based underwater image enhancement. TTA and WBF were used for ensembling. Github

Kaggle: HuBMAP + HPA - Hacking the Human Body [TOP 6%, Bronze Medal]

Problem statement: semantic segmentation of five imbalanced medical tissue types using only 351 high-resolution images from two different train-test sources. Trained Swin Transformer Tiny with an UpperNet decoder, plus CoaT. Used SWA-based ensembling for multiple folds and then TTA across different models. Github

hubmap_kaggle

DDPM: Landscape Generation Using Diffusion Model

Scratch-implemented DDPM using PyTorch. Two variants of U-Net were implemented: a simple CNN-based U-Net and a ViT-based U-Net. Experimented with Classifier-Free Guidance (CFG).

DDPM Project

Carla E2E Autonomous Vehicle Stack

This repository offers a ready-to-use perception and control stack for autonomous vehicles, designed to streamline development and testing in CARLA. Github

carla AV

Object Insertion in Gaussian Splatting

Applied mask-guided 3DGS training for background-extracted models. Utilized Floater-Free MCMC with Bilateral Grid training for large scenes. Performing scalling and positioning for gaussian models. Github

3dgs

3D Structure from Motion (SfM) with Bundle Adjustment

Built a SfM pipeline: SIFT matching, essential matrix–validated pair selection, incremental view addition (PnP/triangulation), bundle adjustment (PyCeres), Numba acceleration, and Open3D visualization. Github

Publications:

Blogs:


For More information check out below links,

Resume | Kaggle | LinkedIn | Twitter | Gmail