Queries regarding Anchor shape and Resizing of image

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi, i currently have a frcnn tlt model (ver 3.0) for object detection. The training images that i currently have are of 1920x1080 resolution but i would like to resize them to 960x540 resolution by setting the output_image_width variable to be 960, output_image_height to be 540 and enable_auto_resize to be True. ( to reduce the size of the model). So currently i have three questions:

  1. for the height and width vairable under the size_hieght_width config, i still set it as 1080 and 1920 respectively am i right?

  2. When calculating for the anchor shapes (under anchor_box_config) to be generated, do i calculate for the resized image(960x540) or for current resolution of image (1920x1080)?

  3. When reducing the size of the model, does it affect the model’s accuracy/performance in detecting the object?

I look forward to your reply, thank you :)

3
  1. No, please set to 960x540 as well. See FasterRCNN — TAO Toolkit 3.22.05 documentation
  2. For the resized image
  3. Usually it will. The smaller size may have a quicker fps but lower accuracy.
4

ok thank you so much for your guidance

6

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.