Introduction to MINERVAS

MINERVAS is a Massive INterior EnviRonments VirtuAl Synthesis system. It aims to facilitate various vision problems by providing a programmable imagery data synthesis platform.

Based on the large-scale (more than 50 million) high-quality (professional artist designed) database of Kujiale.com, MINERVAS provides a way for all users to access and manipulate them for facilitating their data-driven task.

System Pipeline

The pipeline of MINERVAS includes the following parts:

  • Scene Process Stage: In this stage, users can filter scenes by their condition and re-arrange the layout of 3D scenes for domain randomization.
  • Entity Process Stage: This Stage is designed for batch processing entities in the scene. Users can easily use entity-level samplers to randomize attributes of each entity, including furniture (e.g., CAD model, material, transformation), light (e.g., intensity, color), and camera (e.g., camera model, transformation). Modifying the attribute of each object manually is also supported.
  • Render Stage: the system uses the generated scenes to generate 2D renderings with the photo-realistic rendering engine.
  • Pixel Process Stage: In this stage, users can apply pixel-wise processing operations on the imagery data.

Considering the flexibility and ease of use of the system, MINERVAS provides two ways of usage: a user-friendly GUI mode and a flexible programmable mode. In the programmable mode, the pipeline is fully controlled by the Domain-Specific Language (DSL). The DSL of MINERVAS is based on Python programming languages and contains multiple useful built-in functions as we will introduce in the next chapter. Moreover, as the diversity of the data is crucial for learning-based methods, MINERVAS also supports domain randomization both in GUI mode and programmable mode.

In this documentation, we provide a DSL programming guide of MINERVAS along with some examples of vision tasks.

For more information, please visit our project page. For any problem in usage or any suggestion, please feel free to contact us minervas@qunhemail.com.