The ViCoS Cube is a modular demonstration-cell application developed for public and professional presentation of deep-learning-based computer vision in practical scenarios. It was designed as a complete demonstrator that connects camera input, a graphical user interface, application control, and deployable deep-learning models, making research results easier to present outside the laboratory.

Associated links

• Code main app: https://github.com/vicoslab/cube
• Code demo programs: https://github.com/vicoslab/cube-apps
• Code cameras: https://github.com/vicoslab/cube-cameras
• Paper: Demonstration Cell for Showcasing Deep Learning in Practical Applications

Overall system structure

The system is organized as a modular stack with clearly separated responsibilities:

1. Main GUI application provides the operator-facing graphical interface and coordinates the active demo.
2. Camera and acquisition layer captures live image streams from the connected camera setup.
3. Demo-program layer contains individual application modules, each with its own configuration, GUI logic, and model runtime.

This structure makes it possible to reuse the same hardware and interface while switching between very different computer-vision tasks, from industrial inspection to counting, traffic analysis, and robotic manipulation. It also keeps the system maintainable, since new demos can be integrated as separate modules instead of requiring a redesign of the full demonstrator. This is implemented as containirized modules allowing full seperation of depending libraries (Python, PyTorch, CUDA, etc.) for each demo app.

Demo programs currently included

Current demonstrator setup includes the following seven demo programs:

• Few-shot counting, demonstrating counting from only a small number of exemplars.
• Tracking with segmentation, demonstrating segmentation-aware visual tracking.
• Cloth grasp points, a robotic-manipulation demo for grasp-point localization on deformable cloth.
• Tile defects, a surface-defect inspection demo.
• Polyp counting, a counting application for dense biological objects.
• Traffic sign detection, a real-time traffic-sign detection demo.
• Wooden Board classification, a classification-oriented application for wooden boards.

Links to the research methods behind the demos

Where a public research output is available, the demo programs connect to the following underlying methods:

• GeCo2: Generalized-Scale Object Counting with Gradual Query Aggregation for few-shot counting.
• DAM4SAM: A Distractor-Aware Memory (DAM) for Visual Object Tracking with SAM2 for tracking with segmentation.
• CeDiRNet-3DoF: Center Direction Network for Grasping Point Localization on Cloths for cloth grasp point detections.
• SuperSimpleNet: SuperSimpleNet: Unifying Unsupervised and Supervised Learning for Fast and Reliable Surface Defect Detection and No Label Left Behind: A Unified Surface Defect Detection Model for all Supervision Regimes for supervised and unsupervised surface defect detection.
• Polyp counting: A segmentation-based approach for polyp counting in the wild.
• Traffic sign detection: Deep Learning for Large-Scale Traffic-Sign Detection and Recognition