HeinSight4.0 Dataset and Models for Dynamic Monitoring of Chemical Experiments

Datasets:
The HeinSight4.0 dataset comprises 6,031 images of chemical experiments conducted in laboratory settings, primarily involving transparent vessels. It classifies chemical phases into five categories:

• Air:
  o    Empty: Clear air above the liquid level.
  o    Residue: Air contaminated with solid deposits.
• Liquid:
  o    Homogeneous Layer: Clear solutions.
  o    Heterogeneous Layer: Cloudy or turbid liquids.
• Solid:
  o    Solid: Particles or deposits either suspended in liquid or forming a distinct phase.

The images were extracted from videos capturing dynamic chemical processes, enriching the dataset to handle diverse phase behaviors such as dissolution, melting, mixing, settling, and more. Additionally, a vessel dataset containing 6,523 images is included. This dataset incorporates images from the HeinSight3.0 dataset, supplemented with new images of reactors and vessels, to enhance detection across a variety of laboratory equipment and setups.
All images were manually annotated, with bounding boxes marking the regions of chemical phases and their respective classifications. The dataset is split into a 90:10 train/validation.

Models:
Two models were trained on the custom HeinSight4.0 dataset using the YOLOv8 architecture, fine-tuned from pretrained models on the COCO dataset. Included in this release are:
•    Model weights.
•    Training parameters.
•    Evaluation metrics.

Code and Usage:
The models and datasets can be utilized via the associated codebase, available at https://gitlab.com/heingroup/heinsight4.0