SS·

M.Sc. Autonomy Technology — FAU Erlangen–Nürnberg

Exploring how machines see the world in 3D

Researching how machines perceive and understand the 3D world around them.

📍 Erlangen, Germany • Open to research & ML engineering roles
View projects Get in touch
01 / ABOUT

From ADAS to Perception

I'm Sivasubiramaniam Subbiah, a perception researcher based in Erlangen, Germany, currently pursuing an M.Sc. in Autonomy Technology at FAU Erlangen–Nürnberg, where I'm writing my thesis on multimodal aerial datasets. I started out writing path-planning and static-analysis code for production ADAS systems at Visteon, and have since moved toward research: adapting depth-completion and panoptic-segmentation models, and evaluating how LiDAR- and image-based 3D scene understanding compare under matched conditions. I like problems where geometry, sensor physics and learned models have to agree with each other.

  • FocusPerception · Sensor fusion · 3D scene understanding
02 / EXPERIENCE

Experience

Computer Vision Research Intern

Fraunhofer IVI — Ingolstadt, Germany

Jul 2025 – Dec 2025
  • Adapted the DFU depth-completion model from KITTI Raw to SemanticKITTI odometry data, generating dense depth maps for the full dataset.
  • Fed the resulting dense depth into IPFormer, trained the model, and analyzed panoptic-segmentation performance (PQ†).
  • Ran rigorous evaluation and metric verification of the PaSCo model, applying image-based FOV restrictions for fair LiDAR-vs-image comparison on SemanticKITTI / KITTI-360.
  • Assessed PaSCo's zero-shot generalization by cross-evaluating SemanticKITTI checkpoints on KITTI-360, building an on-the-fly WaffleIron feature-extraction pipeline to manage resource constraints.
  • Trained PaSCo under FOV restrictions on SemanticKITTI to enable a direct, equitable comparison against image-based models such as IPFormer.
  • Proposed a latent-space fusion architecture adapting JEPA principles for self-supervised learning on multi-modal data.

ADAS Engineer (Functions)

Visteon Corporation — Chennai, India

Jul 2022 – Sept 2023
  • Worked across Linux, ROS2 and Python alongside C++/C to turn planning theory into running vehicle code.
  • Developed a hybrid A* path-planning algorithm for parking scenarios, combining A* with an Artificial Potential Field (APF) in Python.
  • Benchmarked the algorithm against alternative planners, consistently hitting favorable real-time performance.
  • Integrated the planner into Gazebo and CARLA, improving vehicle map creation via SLAM.
  • Strengthened static-code-analysis coverage against HIS, Coverity and MISRA, eliminating 500+ findings.
  • Built deep familiarity with the VTD simulator, resolving a range of functional use cases.
03 / PROJECTS

Selected projects

Pascal-TriheadNet

Multi-task vision

A single-stage unified perception model sharing one ViT backbone across detection, semantic and instance segmentation.

  • 87.3% semantic mIoU
  • 75.6% detection mAP@50
  • 65.7% instance mask mAP@50

ViTDet-style Simple Feature Pyramid neck feeding FCOS detection, Panoptic-FPN semantic and Mask R-CNN instance heads; unified loader handles heterogeneous XML/PNG annotations; INT8 quantization cut model size 82%.

PyTorch · timm · TorchMetrics · Hydra · W&B
View on GitHub →

Dog Breed Classification

Transfer learning

Fine-grained classifier across 120 dog breeds using pretrained backbones and transfer learning.

  • 82.2% validation accuracy
  • 120 breed classes

Custom PyTorch dataset pipeline with metrics logged to TensorBoard.

PyTorch · PyTorch Lightning · Pandas · TensorBoard
View on GitHub →

Bulldozer Price Prediction

Regression

Random Forest regressor predicting bulldozer resale price on the Kaggle Blue Book for Bulldozers dataset.

  • 0.241 RMSLE
  • ~37.5s training time
Python · scikit-learn · Pandas · Matplotlib
View on GitHub →
04 / SKILLS

Skills & tools

Core

Machine Learning · Deep Learning · Computer Vision · Radar Signal Processing · Motion Planning · Optimization · Robotics · ROS2

Software & tools

Python · Git · Linux · PyTorch · Agisoft Metashape · CloudCompare

Spoken languages

English (advanced) · German (intermediate) · Tamil (native)

05 / EDUCATION

Education

M.Sc. Autonomy Technology

Friedrich-Alexander-Universität Erlangen–Nürnberg

Apr 2024 – present

GPA: 1.7

Coursework: Machine Learning, Deep Learning, Reinforcement Learning, Computer Vision, Radar Processing, Robotics, VSLAM Seminar, Video Compression Lab.

B.E. Electronics & Communication Engineering

College of Engineering, Guindy

Aug 2018 – Jun 2022

GPA: 2.3

Coursework: Image Processing, Signal Processing, C++, C.

06 / CONTACT

Let's talk

Open to research collaborations, perception engineering roles, and interesting problems in 3D vision.

siva14082001@gmail.com LinkedIn GitHub