TAO API - TIFF imagery format for object detection

Accelerated Computing Intelligent Video Analytics TAO Toolkit
,
1

Hi,

For TAO API, do the TAO object detection models support training / inference on GeoTIFF images?

We’ve been following this notebook which indicates JPG and PNG images.

After deploying TAO API to an EKS cluster, I’ve constructed a KITTI PNG format dataset (train, val, images, labels), successfully uploaded it to the TAO server, run TFrecords conversion, and then various TAO YOLOv4 experiment actions (train, export, etc.).

But my question is:

  • Can the same be done with GeoTIFFs? are there TAO API examples I can refer to?
  • What changes / considerations are needed for this vs. the standard JPG and PNG imagery format?

Thank you!

2

Could you please have a try with the tao launcher’s YOLO_v4 notebook(tao_tutorials/notebooks/tao_launcher_starter_kit/yolo_v4 at main · NVIDIA/tao_tutorials · GitHub)? If it can work. The TAO API’s notebook should work as well.

3

Hi, I’m not sure I see references to TIFFs in there. I see:
['.jpg', '.png', '.jpeg', '.ppm']
as image extension options?

4

TAO YOLOv4 does not explicitly support TIFF image format for training or inference. The documentation and discussions primarily mention support for JPG and PNG formats (YOLOv4 - NVIDIA Docs). However, there are a few important points to consider:

  1. TAO YOLOv4 supports 16-bit grayscale images, which is a feature often associated with TIFF files (YOLOv4 - NVIDIA Docs).
  2. If direct TIFF support is not available, a common workaround is to convert TIFF images to a supported format like PNG or JPG before training or inference.