README.md

metadata

pretty_name: GSO-Net
task_categories:
  - object-detection
  - image-classification
tags:
  - computer-vision
  - industrial-vision
  - safety-monitoring
  - sop-compliance
  - benchmark
size_categories:
  - 10K<n<100K
annotations_creators:
  - expert-generated
language:
  - en
license: other

GSO-Net

GSO-Net studies SOP understanding under sparse industrial polling, where procedural status must be inferred from localized operational evidence rather than dense temporal continuity.

GSO-Net is a large-scale benchmark for visual Standard Operating Procedure (SOP) understanding in petrochemical unloading scenarios. It is designed for realistic industrial deployment, where monitoring often relies on sparse round-robin visual polling rather than dense continuous video.

The dataset contains 50,325 independently sampled frames collected from 63 real petrochemical stations. It adopts a hierarchical annotation scheme linking 9 macroscopic procedural steps with 15 microscopic operational states, enabling evaluation of both procedural understanding and localized state grounding.

Highlights

• Real industrial scenario: petrochemical unloading under practical surveillance conditions
• Hierarchical labels: 9 macro steps + 15 micro states
• Sparse-polling setting: built for intermittent observation rather than dense video
• Challenging benchmark factors: small critical targets, long-tailed operational states, cross-site variation, weather and illumination changes
• Engineering-oriented focus: evaluates whether models can support reliable SOP monitoring under deployment constraints

Tasks

Task 1: Joint Step-State Detection

The core task of GSO-Net.

Models jointly predict:

• 15 microscopic operational states
• 9 macroscopic procedural steps

This task evaluates whether a model can ground localized operational evidence and relate it to procedural stages.

Task 2: Frame-Level Step Classification

A diagnostic reference task.

Models classify one of the 9 macroscopic SOP steps from the whole image using global visual evidence only.

This task is intentionally weaker than Task 1 and is included to show the limitation of holistic classification without explicit localized state grounding.

Dataset Scale

• Total images: 50,325
• Stations: 63
• Bounding boxes: 321,432
• Training images: 40,976
• Validation images: 9,349
• Split protocol: strict cross-site split

Why GSO-Net?

Existing industrial benchmarks mainly focus on anomaly detection, object presence, or dense procedural parsing. GSO-Net instead targets hazardous SOP understanding under sparse industrial polling, where procedural status must be inferred from incomplete observations and localized visual evidence.

The benchmark is particularly challenging because:

• many decisive cues are tiny and localized
• operational states are long-tailed
• adjacent procedural stages may look globally similar
• stage boundaries often depend on functional state transitions rather than large scene changes

Data Structure

A typical release includes:

GSO-Net/
├── images/
│   ├── train/
│   └── val/
├── labels/
│   ├── train/
│   ├── val/
├── annotations/
│   ├── instances_train2017/
│   ├── instances_val2017/
├── splits
├── README.md
└── LICENSE