NGC | Catalog
Welcome Guest
CatalogResourcesTAO with Synthetic Data

TAO with Synthetic Data

For downloads and more information, please view on a desktop device.
Logo for TAO with Synthetic Data

Description

Jupyter Notebook for training FasterRCNN with Rendered.ai user generated datasets.

Publisher

Rendedred.ai

Use Case

Other

Framework

Other

Latest Version

obstruction_fasterrcnn_v1

Modified

March 17, 2022

Compressed Size

39.26 KB

TAO with Synthetic Data

The ana_tao resource contains a Jupyter Notebook and supporting configuration files for training the FasterRCNN TAO model using Rendered.ai synthetic imagery data. The notebook is a tutorial that demonstrates the use of the Rendered.ai SDK in conjunction with the NVIDIA TAO toolkit as described in the article, "Using the Rendered.ai Synthetic Data Platform with NVIDIA TAO to Improve Training Data through Closed-Loop ML".

A user can access datasets on Rendered.ai and evaluate a production grade object detection model for various levels of object obstruction. The process helps the user understand how CV models will perform on real-world datasets by enabling modification of data generation configurations, or graphs, to test hypotheses gleaned from viewing inferences. For example, how object detection accuracy is affected by training data with objects at different densities or obstruction levels.

Register now to get started with synthetic data!

Using the ANA_TAO Resource

Extract the ana_tao resource files into a directory, e.g. ana_tao, and make a project based on that top level folder. The project will contain the obstruction_fasterrcnn experiment and will be the root location for Rendered.ai datasets collected when running the experiment.

The first few cells of obstruction_fasterrcnn.ipynb install the python environment dependencies. Optionally, if running locally, the notebook will run with the packages installed for your jupyter kernel.

The notebook concludes with the following results.

Effect of number of objects in a Rendered.ai graph on OD performance

About Rendered.ai

Rendered.ai is a platform-as-a-service for synthetic data that enables data scientists to overcome the costs and challenges of acquiring and using real data for training machine learning and artificial intelligence systems.

View of Rendered.ai GUI

Other versions of ana_tao

Datasets are available for the 'sample_dataset' or 'datasets' versions.

The 'datasets' version contains 3 batches of annoteted images for download

ngc registry resource download-version "isv-ngc-partner/renderedai/ana_tao:datasets"

Dataset Description
200runs_10objectScene 200 images with low object obstruction
200runs_40objectScene 200 images with moderate object obstruction
200runs_70objectScene 200 images with high object obstruction

The 'sample_dataset' version contains 10 images for viewing here in the file browser:

Item Description
sample_dataset/ A sample dataset of 50 objects per scene
10 Object Scene.png Sample of 10 objects per scene
40 Object Scene.png Sample of 40 objects per scene
70 Object Scene.png Sample of 70 objects per scene
100 Object Scene.png Sample of 100 objects per scene