• Hardware Platform (Jetson / GPU) GPU (GTX 1060)
• DeepStream Version 5.0.1
• JetPack Version (valid for Jetson only) N/A
• TensorRT Version 7.0.0
• NVIDIA GPU Driver Version (valid for GPU only) 460.80
Hello,
I have a multilabel classifier model that I’m running as a secondary classifier using nvinferserver plugin. The primary detector is run on the nvinfer plugin.
I noticed that I am getting different output on the classifier labels over a few runs on the same input source. When I disabled the tracker, however, the classifier was able to consistently output the same results. I’ve tried to disable async_mode for the classifier, but results are still inconsistent when the tracker is enabled.
To compare the results between runs, I modified ds-example to dump NvDsObjectMeta.object_id, NvDsObjectMeta.detector_bbox_info, NvDsClassifierMeta.obj_label, and NvDsLabelInfo.result_label from NvDsFrameMeta. The tracking ID, bounding boxes, and object labels are consistently the same, regardless of whether there is a tracker or not. I’ve also separately dumped the GstBuffer data as Base64, and found no differences between runs. Only the classifier labels in NvDsLabelInfo.result_labelare inconsistent when the tracker is on.
I managed to reproduce this issue with deepstream-app using the following configuration:
(modified from samples/configs/deepstream-app-trtis/source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt)
source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt (5.7 KB)
(modified from samples/configs/deepstream-app-trtis/config_infer_secondary_plan_engine_vehicletypes.txt)
config_infer_secondary_plan_engine_vehicletypes.txt (2.2 KB)
I ran this a total of 5 times, each time the NvDsLabelInfo.result_label of some detections are different. When the tracker is disabled in source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt, however, the results between runs become consistent.