Wilke Grosche

Wilke Grosche

German-British Physics graduate (MSc, ETH Zürich) bridging control theory, computer vision, and machine learning.

MultiCam-GT

Published:

Overview

Annotation pipeline for multi-view people tracking datasets, used in the construction of the public SCOUT dataset. Automated tracklet merging reduced manual annotation time by ~90%. (8 months)

Problem

Annotating multi-view pedestrian detection datasets with contiguous overlapping fields of view across 25 cameras requires spatially consistent labelling, temporal coherence of tracks at ~10-minute timescales, and efficient tooling to make the annotation volume tractable.

Features

  • Multi-view annotation: Web-based tool for spatially-aware annotation with cross-view consistency across 25 cameras
  • Automated tracklet merging: Proximity-based tracklet merging and cross-view correspondence to minimise manual effort
  • Trajectory analysis: Support for trajectory analysis at ~10-minute timescales
  • Elevation modelling: Mesh-based ground plane representation for varying terrain
  • Database integration: All transformation and tracking data stored in PostgreSQL

Technical Details

Built with Django and PostgreSQL on the backend, with a JavaScript frontend for interactive annotation. The automated tracklet merging pipeline was the key contribution, reducing the manual annotation burden by approximately 90% and making the construction of the SCOUT dataset feasible.