How to set parameters for SSD sample

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi Guys,

I am unable to understand how to set parameters, specifically scale in the SSD example. This is because the sample runs on training image input size of 1248x384 . Since this is not a square image it is difficult for me to set the parameters for images of different resolution such as 480x480. Additionally, if I use “resnet10” architecture then which layers do the scale parameters refer to. As per documentation, I understand that the ith value in the list refers to the ith layer. Is that the correct understanding?

Please help me out.

Thanks.

2

Hi neophyte1,
In SSD example, output_image_width/height inside training config file is set to 1248x384 because KITTI dataset(mostly 1248x384) is used in the previous step of Jupyter notebook.
The setting in config file should match the width/height of the network input.

Currently for feature extraction architecture in SSD, only “resnet10” and “resnet18” are supported.
For each, there are 6 feature maps.
The aspect_ratios_global is a list of aspect ratios for which anchor boxes are to be generated. This list is valid for all prediction layers as follows.

[[1.0, 2.0, 0.5, 3.0, 0.3333333333333333], [1.0, 2.0, 0.5, 3.0, 0.3333333333333333], [1.0, 2.0, 0.5, 3.0, 0.3333333333333333], [1.0, 2.0, 0.5, 3.0, 0.3333333333333333], [1.0, 2.0, 0.5, 3.0, 0.3333333333333333], [1.0, 2.0, 0.5, 3.0, 0.3333333333333333]]

The aspect_ratios should be a list of lists inside quotation marks. The length of the outer list must be equivalent to the number of feature layers used for anchor box generation.
If it is

"[[1.0,2.0,0.5], [1.0,2.0,0.5], [1.0,2.0,0.5], [1.0,2.0,0.5], [1.0,2.0,0.5], [1.0, 2.0, 0.5, 3.0, 0.33]]"

Then, the i-th layer will have anchor boxes with aspect ratios defined in aspect_ratios[i]. Totally 6 layers.
The last layer has anchor boxes with aspect ratio [1.0, 2.0, 0.5, 3.0, 0.33]

How to calculate Anchor Shape Sizes for Frcnn
3

Hi Morganh,

Thanks for the info. I am unable to still understand which layers are tapped out for SSD. If there are 6 parameters in the list and there are 10 layers (resnet-10), then how do we know which 6 layers out of the 10 layers correspond to the 6 parameters ?

Please let me know.

Thanks.

4

Hi neophyte1,
In TLT SSD, only 6 feature layers are available in resnet10 or resnet18.