Please provide complete information as applicable to your setup.
• Hardware Platform (GPU)
• DeepStream Version: 6.1
• TensorRT Version: 8.2.5
• NVIDIA GPU Driver Version : 11.7
• Issue Type( bugs)
When sgie is connected, the pipeline starts to have the following error at the end of the video. I checked the source code of nvinfer.cpp and found that this is caused by the type mismatch when the sgie node is zoomed and changed. How to solve it?
In this pipeline, I did not use tracker behind pgie, but put sgie directly behind pgie, and added a queue in the middle. I just tried to add the tracker between pgie and sgie, and the above problem did not appear. What is the problem? If I don’t want to use tracker to avoid this problem, is there any other solution?
The entire pipeline can be simply expressed as: uridecodebin->nvstreammux->queue->nvinfer(pgie to detect person face)->queue->nvinfer(sgie to extract face feature)->queue->nvvideoconvert->queue->capsfilter->queue->fakesink; The pipeline takes a local video file as input, and there is no problem during operation, but the above-mentioned problem occurs at the end of the video. The code project I use is reference: GitHub - zhouyuchong/face-recognition-deepstream: Deepstream app use retinaface and arcface for face recognition.
please refer to the NV official sample deepstream-faciallandmark-app, which works fine and has similar pipeline.
I have found a solution, please refer to: convert_batch_and_push_to_input_thread:<sgie> error: NvBufSurfTransform failed with error -3 while converting buffer · Issue #24 · zhouyuchong/face-recognition-deepstream · GitHub
Thanks for your sharing, please help to to reproduce that "obj.left < 0 " issue, then we will fix the bug, is it related to the source video? could you share the source video? thanks.
This problem is relatively easy to trigger, as long as pgie sets symmetric-padding=1 and maintain-aspect-ratio=1, and there is sgie in the pipeline, at this time, if the target is detected to be on the edge in the video, the problem is particularly easy to be triggered . In addition, the video I use is currently private data, restricted by the company’s confidentiality, so it is not convenient to provide, sorry.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.
