The two test results are different in secondary gie stage

[Mike_deng]

• Hardware Platform ：T4
• DeepStream Version ：6.1
• JetPack Version (valid for Jetson only)
• TensorRT Version ： 8.2.5

I run two stages model with deepstream, second model use primary’s output.
after run two times, wo found the resul in primary are same, but the result in secondary-gie0 are different.

the config file:

[primary-gie]
enable=1
gpu-id=0
model-engine-file=1trt
labelfile-path=labels.txt
batch-size =1
#Required by the app for OSD, not a plugin property
bbox-border-color0=1;0;0;1
bbox-border-color1=0;1;1;1
bbox-border-color2=0;0;1;1
bbox-border-color3=0;1;0;1
interval=0
gie-unique-id=1
nvbuf-memory-type=2
config-file=config_infer_primary.txt

[secondary-gie0]
enable=1
model-engine-file=4.trt
gpu-id=0
batch-size=1
gie-unique-id=4
operate-on-gie-id=1
operate-on-class-ids=0;
labelfile-path=labels.txt
config-file=config_infer_secondary.txt

[fanzh]

need more information, which deepstream sample are you testing? did you modify the code, models? how did you get the results above?

[Mike_deng]

yes, i change some code,but i don’t change the inference parts, I will confirm in original codes.

the log was print in gst-dsexample:
for (l_obj = frame_meta->obj_meta_list; l_obj != NULL; l_obj = l_obj->next)
{
obj_meta = (NvDsObjectMeta *)(l_obj->data);

    gint src_left = GST_ROUND_UP_2((unsigned int)obj_meta->rect_params.left);
    gint src_top = GST_ROUND_UP_2((unsigned int)obj_meta->rect_params.top);
    gint src_width = GST_ROUND_DOWN_2((unsigned int)obj_meta->rect_params.width);
    gint src_height = GST_ROUND_DOWN_2((unsigned int)obj_meta->rect_params.height);
   std::cout<<....

}

[fanzh]

which deepstream sample are you testing? did you modify the models? what is the whole media pipeline? can you use deesptream native sample to reproduce this issue? deepstream-test2 supports pgie+sgie.

[Mike_deng]

I found same problem
in nvcr.io/nvidia/deepstream :6.1-triton after two time in source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt ,
I just modify the source4 to source0, and add print in ds-example.txt
source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt (5.8 KB)

[fanzh]

Using the same cfg, code patch with yours, I can’t reproduce this issue on T4 deepstream6.1, here are code, test results.

code.txt (527 Bytes)
log1.txt (717 Bytes)
log2.txt (662 Bytes)

[Mike_deng]

thanks very much, the problem was fixed , the reason :classid is not assigned . it is a random, the box was dropped when the value>num-detected-classes.