Back to projects
Model DevelopmentFeature Extraction

ViTDet Object Detection

A practical reconstruction of a Vision Transformer detection pipeline using Detectron2 and ViTDet concepts.

Role
Model implementation and research engineering
Focus
Model Development / Feature Extraction
Stack
PyTorch, Detectron2, ViTDet

Context

The shape of the problem

This was research-oriented engineering: understanding how transformer-based object detection works by rebuilding the moving pieces rather than only consuming a high-level API. The value was in the implementation detail and evaluation discipline.

Build

What I actually built

I adapted ViTDet-style components with Detectron2, worked through COCO-style data handling and evaluation, and connected the detection stack to familiar Mask R-CNN and Cascade R-CNN techniques. The project sharpened the bridge between papers, frameworks, and working code.

Engineering

Key decisions

01

Understanding and adapting transformer-based architectures for object detection

Adapted ViTDet architecture using PyTorch, gaining practical knowledge of vision transformers

02

Creating a functional detection pipeline combining multiple components like backbone networks, feature pyramid networks, and region proposal networks

Implemented a detection pipeline incorporating Cascade R-CNN and Mask R-CNN techniques

03

Setting up accurate evaluation metrics using COCO standards

Developed comprehensive data handling and basic evaluation workflows

Outcome

What changed

  1. 01Achieved a fully functional detection pipeline demonstrating transformer effectiveness in vision tasks
  2. 02Successfully evaluated using COCO metrics, establishing foundational skills in state-of-the-art object detection techniques

Stack

Tools and systems

PyTorchDetectron2ViTDetComputer Vision

I build delightful systems.

© 2026 Kristian Ernst.