TAO vs isaac_ros_dnn_inference encoder param mismatch

[automech-rb]

For models exported by tao toolkit v4.0 (yolov v4 for my case): It seems like encoder_image_mean and encoder_image_stddev needs to be set to 0 and 1/255 respectively to get any proper inference (thereby nullifying the normalization). When normalizing image between 0 to 1, TAO exported models never output correct predictions.

Is there any way to turn off normalization by a parameter for encoder node?

[jaiveersinghNV]

As you've observed, the encoder_image_mean and encoder_image_stddev parameters can be set to 0 and 1/255 respectively to effectively turn the normalization step into a no-op.

We will look into adding a more explicit flag for disabling normalization.