All projects

Counting Machine Washer Parts

An industrial challenge to count highly occluded parts from imagery.

Dec 2021 Machine Learning

A computer-vision pipeline to count densely-packed, highly-occluded industrial washer parts.

  • Benchmarked image-thresholding methods: Otsu thresholding, Gaussian filtering, Sobel filtering and Hough circle transforms.
  • Extended density-map-based FamNet with a novel mismatch-loss component.
  • Achieved 1.96 MAE (a 90% reduction from the image-processing baseline) on the challenge dataset.
Density EstimationFamNetObject Counting