"If data is oil, Machine Learning is oil refinery”-- Open Source Image Sampler

Keeping this thought in the mind, oil needs to be mined and made available.

One of the pain point and challenge for machine learning engineers is lack of appropriate image data for their specific projects and often such images are not available freely or not in the enough quantity. The idea of this project is to capture images of any object/product/item at home with speed. 

To get best quality and appropriate image samples the paper concludes following:

• Azimuth - Object should be revolved 360 degree at least 24 times in equal angle. This means 360/24 = 15 degree.
• Elevation - In a circular fashion object should be captured in four verticals i.e. 90 degree / 4 = 22.5 degree.

Open Source Repo Contains: 

1. Word document that gives high level details about the sampler.
2. Mechanical Design: CAD file of the sampler design along with BOM
3. Source code to run controller and DSLR controller software to capture pictures.

Working Flow:

Following actions and events occur during sampling:

1. Place the object on the revolving table top,
2. Enter the barcode so that the respective folder will be created with the name as ‘barcode #’. Else enter manual name.
3. Follow the on-screen messages to let table top to start revolving and Cameras will start capturing images. There are total 5 cameras capturing images in tandem.
4. Once 360 degree movement is over, the table top stops.
5. To capture another object jump to step 1.

Physical Images Of Sampler

GitHub Repo

Future Work

We welcome participation from developers community. Following are the features we look forward to:

1. Enhancing physical design.
2. Data Augmentation at runtime.
3. Full DSLR control to the users.

Contributors

Saket Deshmukh (Saaket.deshmukh@gmail.com) Milind Deore (tomdeore@gmail.com)