There is a problem that the prediction results are missing on tao yolo v4 tiny

[kysim]

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) : Jetson AGX Xavier
• DeepStream Version : 5.1
• JetPack Version (valid for Jetson only) : 4.6 (cuda and trt version match jetpack 4.5.1)
• TensorRT Version : 7.1.3
• Issue Type( questions, new requirements, bugs) : questions

I tested detection performed through deepstream + tao yolo v4 tiny, but there is a problem that detection is omitted at the edge of the screen.

• DetectNet :
  • tao / deepstream-app : working
• YOLO :
  • tao inference / nvinfer - custom_parser : working
  • deepstream-app result : detection is omitted at the edge of the screen.
    
    
    
    

These are using config files.
pgie_config_yolo_v4.txt (766 Bytes)
source.txt (3.1 KB)

Why is this problem happening and how can it be solved?

thanks.

[AastaLLL]

Hi,

First, it’s recommended to upgrade your software to the latest JetPack 4.6.1 and Deepstream 6.0.1.

There is a configuration to control the ROI output offset.
The default value is set to 200, indicating the bounding box needs to be away from the top/bottom 200 pixels.

Would you mind updating the following two configures to see if it work?
https://docs.nvidia.com/metropolis/deepstream/6.0/dev-guide/text/DS_plugin_gst-nvinfer.html#id3

[class-attrs-all]
pre-cluster-threshold=0.5
roi-top-offset=0
roi-bottom-offset=0

Thanks.

[kysim]

Thank you for your reply!

I will check and reply again.

[labar90]

I have the same behaviour with DS 6.0.1 and the ROI properties with 0px. Detections are omitted too.

[AastaLLL]

Hi, labar90

Have you applied the flag shared above?
If not, would you mind giving it a try?

There is a problem that the prediction results are missing on tao yolo v4 tiny

Hi, First, it’s recommended to upgrade your software to the latest JetPack 4.6.1 and Deepstream 6.0.1. There is a configuration to control the ROI output offset. The default value is set to 200, indicating the bounding box needs to be away from the top/bottom 200 pixels. Would you mind updating the following two configures to see if it work? https://docs.nvidia.com/metropolis/deepstream/6.0/dev-guide/text/DS_plugin_gst-nvinfer.html#id3 [class-attrs-all] pre-cluster-threshold=0.5 roi-top-of…

Thanks.

[kysim]

I tested it by 1) upgrading jetpack and deepsteam version and 2) adding the roi-{top, bottom}-offset option. The problem was not solved.

The problem has been resolved by applying the answers and further modifying the custom parser!

GitHub - NVIDIA-AI-IOT/deepstream_tao_apps: Sample apps to demonstrate how to...

deepstream_tao_apps/post_processor/nvdsinfer_custombboxparser_tao.cpp
 (NvDsInferParseCustomBatchedNMSTLT)

[line 168]
object.height = CLIP(p_bboxes[4*i+3] * networkInfo.height, 0, networkInfo.height - 1) - object.top;
=>
object.height = CLIP(p_bboxes[4*i+3] * networkInfo.height, 0, networkInfo.height - 4) - object.top;

----------
networkInfo.height {- 3 ~ -1} : omitted
networkInfo.height {... ~ - 4} : not omitted

Thanks!

[AastaLLL]

Thanks for sharing this.
Good to know you fix this now.