Reproduce “Automated Counting of Stacked Objects in Industrial Inspection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Corentin Dumery, Noa Etté, Aoxiang Fan, Ren Li, Jingyi Xu, Hieu Le, Pascal Fua

Researchers from EPFL developed 3DC, a multi-view computer vision method for automated counting of heavily occluded, irregularly stacked 3D objects in industrial settings. This approach decomposes the task into 3D stack volume estimation and volumetric occupancy ratio prediction, significantly outperforming human estimation and traditional baselines with a Normalized Absolute Error (NAE) of 0.36 and a Symmetric Mean Absolute Percentage Error (sMAPE) of 53.31 on real-world industrial inspection data.

Key insights

  • Decomposing the complex task of counting occluded objects into distinct geometric (volume) and density (occupancy ratio) estimations is more effective than direct count prediction by neural networks.
  • Integrating pre-trained vision models for feature extraction and depth estimation into a specialized architecture helps bridge the domain gap between synthetic training data and real-world performance.
  • The formal derivation of object count based on volumetric principles provides a robust theoretical foundation for the proposed decomposition strategy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.