Preprocess objects before sgie

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson Nano)
• DeepStream Version 6.0
• JetPack Version 4.6.4-b39
• TensorRT Version
• Issue Type( Modifying detected objects before sgie. I have deepstream pipeline for gender detection, first I use pgie to detect face and an sgie that classifies them into male and female. I want to change it like this, for every detected object i want to get its crop make it a little bigger and then run sgie on the crop not on original one. Is there any i can do that, one solution was to get bbox of detected object and then modify it as per need and add it to metadata so that sgie infers on this crop now but i am not sure if this achievable. I need to do this to achieve better accuracy)

There may be a simple method to meet your needs. You can enlarge the image through the nvstreammux. Then the images cropped will be enlarged too.

I did not want to just enlarge the crop by resizing or making the input frame bigger like you suggested. What I really wanted was this, once the model has made predictions, I wanted to increase size of bbox by some percentage and then extract this bigger crop and make sgie infer on this bigger crop not on original one(As the face detector predicts concise size of face I want to include some part of hair and other facial features by just making the crop bigger). But never mind there was a simple solution for this too I just changed bbox parameters in metadata after pgie and i got what i wanted.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

Yes. If you want to increase the dims of the bbox, you can only change the value of bbox by yourself.

Hey I have one question. I have same pipeline as the mentioned in the post, I am running it on jetson nano which has deepstream version of 6.0 and I am running it on a GPU system which has deepstream verison 6.3 but there is considerable mismatch in accuracy of gender classifications why is that? Accuracy on nano is much lower.

The TensorRT, which deepstream relies on for advanced versions, will have a higher version. The version upgrade may improve the accuracy of inferencing. In terms of improving accuracy, we also have relevant FAQ for referring to.

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