Scale problem when use sgie

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

hi,
i am working on jetson nx with deepstream 6.0.
i am using 608608 shape yolo model as pgie, and fd_lpd model as sgie1 to detect face,
and ir50 model 112112 shape as sgie2 to recognition face, ir50 model’s net-scale-factor=0.0078125.
for fd_lpd model and ir50 model, i have setup for both:
input-object-min-width=64
input-object-min-height=64

now problem is:
deepstream-app.c pipline works well when i test in room, but happens segment error always when i test out of door, i think maybe it is a scale problem for small objects but i can not find solution.
pls help check where is problems?

thank you very much!

#3

Can you run gdb to check the segmant fault?

#4

sorry i don’t know that, how to run that?
pls give some advice how to increase sgie nvifer buffer, too.

thank you very much!

#5

hi,
i push back person name of chinese language from json file to change label of sgie object “face” by the line:
strcpy(obj->text_params.display_text, str.c_str());

and deepstream support chinese language?
i found sometimes it is error like:

label error
label error734×462 65 KB

#6

forget that and i didn’t push back correct label for multi objects to get buffer blocked.
and no problem for chinese language, too.
i will test again.

thank you

#7

Glad to know you fixed your issue. Can you share more on how to fix it?

#8

hi,
i tested again outdoor by far away object, and same problems again.
everything ok in room but error segment outdoor…
pls give an advice how to get double gst buffer for inference?

thank you very much!

#9

Please submit new topic if it is one new issue. Thanks.

#10

it is same issue.

thanks

#11

another possible is sending too many frames to cause memory be blocked.
and also pls advice how to clear buffer after inference.

thank you very much!

#12

Please submit new topic if it is one new issue. Thanks.

#13

forget that and scale problem confirmed, works now.
anyway thank you gdb advice.
deepstream is good deepstream for ever, but code worker is stupid sometimes…

thank you