Detectnet - Mounted frames

agarcesc7 · October 27, 2021, 1:59am

Hi,

I am doing a training test with images of types of apples, in the response of the system I am getting the squares mounted with the label.
How can I improve that answer?

AastaLLL · October 27, 2021, 6:16am

Hi,

Do you want to merge the bounding box?
If yes, you can check the NMS algorithm:

github.com

dusty-nv/jetson-inference/blob/master/c/detectNet.cpp#L843


      
          	

          	const float scale_x = float(width) / float(GetInputWidth());

          	const float scale_y = float(height) / float(GetInputHeight());

          

          	const float bbox_norm = 35.0f;  // https://github.com/NVIDIA-AI-IOT/tao-toolkit-triton-apps/blob/edd383cb2e4c7d18ee95ddbf9fdcf4db7803bb6e/tao_triton/python/postprocessing/detectnet_processor.py#L78

          	const float offset = 0.5f;      // https://github.com/NVIDIA-AI-IOT/tao-toolkit-triton-apps/blob/edd383cb2e4c7d18ee95ddbf9fdcf4db7803bb6e/tao_triton/python/postprocessing/detectnet_processor.py#L79

          	

          #ifdef DEBUG_CLUSTERING	

          	LogDebug(LOG_TRT "input width %u height %u\n", GetInputWidth(), GetInputHeight());

          	LogDebug(LOG_TRT "cells x %i  y %i\n", cells_x, cells_y);

          	LogDebug(LOG_TRT "cell width %f  height %f\n", cell_width, cell_height);

          	LogDebug(LOG_TRT "scale x %f  y %f\n", scale_x, scale_y);

          #endif

          

          	for( uint32_t c=0; c < mNumClasses; c++ )   // c = current object class

          	{

          		for( uint32_t y=0; y < cells_y; y++ )

          		{

          			for( uint32_t x=0; x < cells_x; x++)

          			{

          				const float confidence = conf[c * numCells + y * cells_x + x];

Thanks.

agarcesc7 · October 27, 2021, 12:43pm

Hello,

Yes, I want to stabilize the detection and that it recognizes with 1 single box and not 3 as shown in the attached image.

In what exact part of the code should it intervene?

dusty_nv · October 27, 2021, 1:52pm

Hi @agarcesc7, try decreasing the clustering threshold on this line in jetson-inference/c/detectNet.h:

https://github.com/dusty-nv/jetson-inference/blob/9bf55495d6a4ff1946dddbb0e81102d2d05f4952/c/detectNet.h#L483

By default the threshold is .75 (meaning the overlapping area needs to be >= 75% of the total area of the bounding box to be merged). So if you decrease that, it will more aggressively merge the bounding boxes.

After you change it, re-compile / re-install the code:

cd /path/to/your/jetson-inference/build
cmake ../
make
sudo make install

agarcesc7 · November 10, 2021, 4:17am

Hi Dusty. Thank you.

It worked! Is there an option to leave only the marked border and the fill without color?

dusty_nv · November 10, 2021, 3:45pm

Hi @agarcesc7, sorry there isn’t the option for that. You could disable the overlay entirely and draw your own bounding boxes (i.e. with openCV or similar)

system · November 24, 2021, 3:45pm

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