Sgie's bbox goes beyond the pgie

[lewis_bo]

**• Hardware Platform (Jetson / GPU) Jetson orin nx
**• DeepStream Version: 6.2
**• JetPack Version (valid for Jetson only) 5.1.1
**• TensorRT Version8.2.2

When I run two detection models of yolov5n, one pgie(vehicle) and one sgie(tail), the sgie’s bbox goes beyond the pgie,
the srcWidth=1920 and srcHeight=536.

The result is shown below

And my pipline :

[fanzh]

1. is the green circle ROI?
2. can you use this simple sample to test your models? this sample supports two back to back detectors.

[lewis_bo]

Two red ROI, pgie(vehicle) → sgie(tail).
I think the tail detector’s bbox should be inside vehicle0’s bbox, but it isn’t

[lewis_bo]

2. I try it， but it not correct.
   The tail relative to the top of the vehicle is 402, but the vehicle is only 395，beyond that.

[fanzh]

from the highlight data in the picture, 395 should be the vehicle’s height, the vehicle’ bottom should be 134+395.
why dose this blue car have no tail part?

[lewis_bo]

The tail part of the blue car is the highlighted part of the oval.
Because the top of tail is 402 and the top of vehicle is 134, 402+134=536 . It overlaps the border of picsrc

[fanzh]

the vehicle’ bottom is 134+395 = 529, the tail 's bottom is 402+80 = 482. the tail is still inside of vehicle rect because 482 is smaller than 529.

[lewis_bo]

I think the top of tail 402 is relative to 134， not relative to picture. so the tail’ bottom is 134+402+80 = 616.

[fanzh]

could you provide the tail model and source video? Thanks, I will have a try, you can share with forum private email.

[lewis_bo]

sure, just only tail model, not need vehicle model ?

[fanzh]

Thank you, please the whole project if you can.

[lewis_bo]

I send the demo refer simple sample. Please check ，thanks

[fanzh]

please try this fix:
fix.txt (927 Bytes)

1. add this code to /opt/nvidia/deepstream/deepstream/sources/gst-plugins/gst-nvinfer/gstnvinfer_meta_utils.cpp.
2. In gst-nvinfer directory, rebuild libnvdsgst_infer.so, then replace /opt/nvidia/deepstream/deepstream/lib/gst-plugins/libnvdsgst_infer.so with the new so.

[lewis_bo]

Thanks for your solution, I changed it a little bit earlier, like this

I just want to know the real cause of this problem, because the tail’ top is out of vehicle’ height , it mean the center coordinate（x,y） is out of source, But this does not happen in yolov5.

[lewis_bo]

I review the preprocess’ source of about “maintain-aspect-ratio=1, symmetric-padding=1”, it right.
But I have a doubt about the part of RGB

when the infer start(gst_nvinfer_start),

It always is NVBUF_COLOR_FORMAT_RGBA when I use VIC to process my picture
But my model is RGB

[fanzh]

1. did you try the fix code? can it fix the original issue?
2. let’s focus on the original issue, if meeting a new issue, please open a new topic, thanks!

[lewis_bo]

Yes, it’s fixed. Thanks for your help
I will open a new topic about new issue