Replicator for Synthetic Data Generation

Status
Open

We estimate that the final, edited video will be 10-15 minutes in duration.

Goal: Show how to generate datasets for AI training (Computer Vision).

  • Can be Bounding Box, Segmentation, Depth etc.

Outline/Video Flow:

  1. Introduce yourself and the goal of the video/what will viewers learn.

      2. Please include:

    1. What synthetic data generation (SDG) is and why is it important? (e.g., filling gaps where real data is expensive)
  2. How do you setup this project? What is the structure?
    • How do you set up a scene and a Replicator script?
    • Explain the folder structure and key files.
    • Briefly explain the role of each key file.
  3. How do you implement the randomization and synthetic data generation?
    • How do you randomize environment variables, such as lighting, materials, object placement? Why is this randomization important?
    • How do you define annotations (i.e. bounding boxes, segmentation masks, keypoints)?
    • How do you export the dataset?
    • Demo: Generate a small dataset of x, y with randomized lighting and show exported JSON + images.
  4. How do would you use this dataset for AI computer vision training?
    • How would you train a model, i.e. using RoboFlow or custom models like FasterRCNN, Yolo.. etc using this dataset?
  • Takeaway: Understand how to prototype synthetic datasets for training models.
  • Additional info:
    • Best to build on top of existing tutorials from NVIDIA docs and reference them
    • Please make a few suggestions on synthetic data topics