Composite Image Generation Block for Edge Impulse

This is a transformation block that generates synthetic data for your Edge Impulse project by compositing background images with randomly placed object images (with transparency). It uses the method described in this tutorial: https://github.com/edgeimpulse/conveyor-counting-data-synthesis-demo. To add new object images and background images store them in the composites folder and use the following command to upload to the block:

$ edge-impulse-blocks push

from within this folder

How to run (Edge Impulse)

Using the Synthetic Data feature (professional & enterprise)

The Studio now has a feature for interacting with Synthetic Data directly from the Studio; and this composite image generation block is available there.

1. Go to a Professional or Enterprise project, choose Data acquisition > Synthetic data.
2. Select the 'Composite Image Generator' block, fill in your parameters, and hit Generate data.
3. Your synthetic data will show on the right side, for quick evaluation of your parameters.

Understanding the Parameters:

The following parameters are used in the parameters.json file to configure the UI. There are a number of options for image processing and transformation tasks:

Composite Parameters

• Composite Directory

  • Description: The folder where the source composite images are found. This folder should contain background and object subfolders.
  • Type: string
  • Default Value: composites
  • Parameter: composite-dir

• Labels

  • Description: A comma-separated list of objects to generate images for. Set to all to generate images for all objects.
  • Type: string
  • Default Value: all
  • Parameter: labels

• Number of Images

  • Description: The number of images to generate.
  • Type: int
  • Default Value: 10
  • Parameter: images

• Object Area

  • Description: The coordinates (x1, y1, x2, y2) of the valid area to place objects in the composite image, or -1 for the whole image.
  • Type: string
  • Default Value: -1
  • Parameter: object-area

• Maximum Number of Objects

  • Description: The maximum number of objects to generate in each image.
  • Type: int
  • Default Value: 5
  • Parameter: objects

• Allow Overlap

  • Description: Whether objects are allowed to overlap.
  • Type: boolean
  • Default Value: false
  • Parameter: allow-overlap

• Allow Rotate

  • Description: Whether to apply random rotation to objects.
  • Type: boolean
  • Default Value: true
  • Parameter: allow-rotate

Image Effects

• Apply Motion Blur?

  • Description: Whether to apply blur to objects to simulate motion.
  • Type: boolean
  • Default Value: false
  • Parameter: apply-motion-blur

• Motion Blur Direction

  • Description: The direction to apply blur to objects to simulate motion (-1 for random).
  • Type: int
  • Default Value: -90
  • Parameter: motion-blur-direction
  • Show If: apply-motion-blur is true

• Apply Fisheye Lens Effect

  • Description: Whether to apply fisheye lens effect to the final images.
  • Type: boolean
  • Default Value: false
  • Parameter: apply-fisheye

• Apply Fisheye Lens Effect to All Layers

  • Description: Whether to apply fisheye lens effect to all layers or just to the objects (in case your background images are already fisheyed).
  • Type: boolean
  • Default Value: true
  • Parameter: apply-fisheye-all-layers
  • Show If: apply-fisheye is true

• Fisheye Lens Effect Strength

  • Description: The strength of the fisheye lens effect.
  • Type: float
  • Default Value: 0.5
  • Parameter: fisheye-strength
  • Show If: apply-fisheye is true

• Crop to Fisheye?

  • Description: Whether to crop the image to remove black borders after applying fisheye lens effect.
  • Type: boolean
  • Default Value: true
  • Parameter: crop-fisheye
  • Show If: apply-fisheye is true

Upload Parameters

• Upload to Category
  • Description: The category to upload data to in Edge Impulse.
  • Type: select
  • Default Value: split
  • Parameter: upload-category
  • Options:
    • split: Split 80/20 between training and testing
    • training: Training
    • testing: Testing

Customizing this repository (enterprise only)

You can modify this repository and push it as a new custom Synthetic Data transformation block.

1. Install the Edge Impulse CLI.

2. Open a command prompt or terminal, and navigate to this folder.

3. Create a new transformation block:

   $ edge-impulse-blocks init
   
   ? Choose a type of block: Transformation block
   ? Choose an option: Create a new block
   ? Enter the name of your block: Custom Composite Image Generator
   ? Enter the description of your block: Generate images by compositing background and object images.
   ? What type of data does this block operate on? Standalone (runs the container, but no files / data items passed in)
   ? Which buckets do you want to mount into this block (will be mounted under /mnt/s3fs/BUCKET_NAME, you can change these mount points in the Studio)?
   ? Would you like to download and load the example repository? no
   

4. Push the block:

   $ edge-impulse-blocks push
   

5. To show this block in Synthetic Data:

   1. Go to your Edge Impulse organization.

   2. Choose Custom blocks > Transformation, find your new block, click the three dots, and select Edit transformation block.

   3. Enable 'Show in Synthetic data':

      

   4. Save the block.

6. You can now access your modified block from Data acquisition > Synthetic data on any enterprise project.

How to run (locally)

1. Install the dependencies:

   pip3 install -r requirements.txt
   

2. Run the generate script:

   EI_PROJECT_API_KEY=ei_034... python3 transform.py --composite-dir composites --labels "all" --images 3 --objects 5 --allow-overlap 1 --allow-rotate 1 --allow-motion-blur 1 --motion-blur-direction -1 --object-area "0,0,100,100"
   

   Replace replace ei_034... with your Edge Impulse API Key (Dashboard > Keys).

3. Output is stored both in output/ and uploaded to your Edge Impulse project.

--synthetic-data-job-id argument / x-synthetic-data-job-id header

If you want to build your own custom Synthetic Data block, you'll need to parse the (optional) --synthetic-data-job-id argument. When uploading data to the ingestion service you need to then pass the value from this argument to the x-synthetic-data-job-id header. transform.py implements this. This is required so we know which job generated what data, and is used to render the UI on the Synthetic Data page.