DetectNet v2 dataset augmentation opposite zoom min/max

[R2P2]

• Network Type: Detectnet_v2
• TLT Version: 4.0.1 (docker image 4.0.1-tf1.15.5)

We are training a DetectNet_v2 on our own dataset. Our dataset has a known bias towards a specific object size (~100px w/h). In the training spec, we set the zoom augmentation to prevent the network from biasing towards a target size:

• zoom_min: 0.5
• zoom_max: 5.0

However, after continuing to have issues detecting larger objects, we visualized some of the images/results with tensorboard since we do not perform data augmentation offline. All training sessions show images zoomed way out (up to 5x). Per Experiment Spec File: meaning of zoom_min and zoom_max, zoom_max is supposed to be zooming in (i.e. image gets larger), however all of our images in the tensorboard are shown to be zoomed out instead (zoom_min).

Below, is one of the tensorboard images that shows the training data being zoomed out by a factor of approximately 4 despite the zoom specification limiting the zoom out to 0.5 (there is translation as well but this is expected). Is the decription in the link above incorrect or is there a bug flipping the zoom_min and zoom_max augmentation properties?

[Morganh]

A zooming ratio of 1.0 will not affect the image, while values higher than 1 will result in ‘zooming out’ (image gets rendered smaller than the canvas), and vice versa for values below 1.0.
That link’s description is needed to modified. I already modify it.
The source code is in tao_tensorflow1_backend/nvidia_tao_tf1/core/processors/augment/spatial.py at 2ec95cbbe0d74d6a180ea6e989f64d2d97d97712 · NVIDIA/tao_tensorflow1_backend · GitHub. You can also refer to it.

[R2P2]

Thanks for confirming! I would recommend clarifying this on https://docs.nvidia.com/tao/archive/5.3.0/text/object_detection/detectnet_v2.html#augmentation-module as well since (in my opinion), most people assume a larger zoom means zooming “in” (i.e. making the source image larger).