Deepstream automatically crops the input frame using bbox from pgie and run sgie on this cropped frame. I don’t want it to behave like that. After pgie, I have landmarks output and I will use them to do warpaffine on full frame, this will output a cropped frame without touching the bbox, And I want sgie to run on this cropped frame instead the one cropped using bbox.
Please tell me if this is possible or not. If possible, can I have some suggestions on how to do it properly? Thank you.
Could you check deepstream_tao_apps/apps/tlt_others/deepstream-faciallandmark-app at release/tlt3.0 · NVIDIA-AI-IOT/deepstream_tao_apps · GitHub to see if any help?
Sorry, I don’t see my case there. I want to discuss a little about SGIE.
In official document :
Secondary mode: Operates on objects added in the meta by upstream components
Normally deepstream automatically uses the bbox from PGIE’s output to crop the full frame into a region and SGIE will run on this region. Do I understand it correctly?
i think in this post can solve your problem
Correct, so the sgie will perform on the modified bbox if you change the bbox on a probe which installed on the sink pad of sgie, that’s the reason why I give you above sample for your reference
My problem can be illustrated like below:
Hope you can understand it and give me some help. Thank you.
Thank you but it seems like ppl still get stuck there. I’ll keep an eye on it.
Ok, I got, then you can refer the topic @PhongNT mentioned, currently you can do it like this.
BTW, we will release a new gst plugin for customized preprocess in later release, then you can do this via that plugin.
