Rapid Vision

Rapid Vision (rv) is a lightweight toolset for generating labeled synthetic image datasets with just a few lines of code.

Advantages

1. Photorealistic results using Cycles rendering engine
2. Simple and clean Python API
3. Automatic segmentation labeling creation
4. Completely open source
5. Seamless integration with Blender's rich procedural toolset.

Getting started

1. Install dependencies

   Install Go and Blender.

2. Install the rv tool

   go install github.com/Rapid-Vision/rv@latest

3. Create scene script

   import rv
   
   class BasicScene(rv.Scene):
       def generate(self):
           self.get_world().set_params(sun_intensity=0.03)
           cube = (
               self.create_cube()
               .set_location((1, 0, 0.5))
               .set_scale(0.5)
               .set_tags("cube")
           )
           sphere = (
               self.create_sphere()
               .set_location((-1, 0, 1))
               .set_shading("smooth")
               .set_tags("sphere")
           )
           plane = self.create_plane(size=1000)
           empty = self.create_empty().set_location((0, 0, 1))
   
           cam = (
               self.get_camera()
               .set_location((7, 7, 3))
               .point_at(empty)
           )

4. Preview the scene

   rv preview scene.py

   Don't close the preview window yet.

5. Randomize the scene

   See how the preview changes on each file save.

   import rv
   from random import uniform 
   
   class BasicScene(rv.Scene):
       def generate(self):
           self.get_world().set_params(sun_intensity=0.03)
           cube = (
               self.create_cube()
               .set_location((1, 0, 0.5))
               .set_scale(0.5)
               .set_tags("cube")
           )
           cube.rotate_around_axis(            
               rv.mathutils.Vector((0, 0, 1)), 
               uniform(0, 360),                
           )                                   
           sphere = (
               self.create_sphere()
               .set_location((-1, 0, 1))
               .set_shading("smooth")
               .set_tags("sphere")
           )
           plane = self.create_plane(size=1000)
           empty = self.create_empty().set_location((0, 0, 1))
   
           cam = (
               self.get_camera()
               .set_location((7, 7, 3))
               .point_at(empty)
           )

6. Render the final result

   rv render scene.py

7. See the resulting dataset

   Resulting dataset has following directory structure:

   out/
   └── 1/
       └── 424702d4-b28e-4082-b5e5-5499f9a49065/
           ├── _meta.json
           ├── Image.png
           ├── PreviewIndexOB.png
           ├── IndexOB.png
           ├── Alpha.png
           └── Noisy Image.png

   Name	Description
   out	Root directroy containing results of all runs
   1	Directory containing results of a single rendering run. Number increases sequentially
   424702d4-b28e-4082-b5e5-5499f9a49065	Directory with resulting image, labeling and additional render passes
   _meta.json	Labeling information
   Image.png	Resulting image
   PreviewIndexOB.png	Preview for the segmentation masks
   IndexOB.png	Segmentation masks. It is a 16-bit black and with image with each pixel corresponding to the index field of each object in the _meta.json file.
   Other	Other rendering passes that can be useful for computer vision tasks and can be additionaly enabled by the Scene.set_passes method.

8. Use rv for generating your next synthetic dataset

   For more information view the Features Overview and API reference.