Problem

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) Jetson Nano
• DeepStream Version 5.1
• JetPack Version (valid for Jetson only) 4.5.1
• TensorRT Version 7.1.3+cuda10.2
• Issue Type( questions, new requirements, bugs) Question

Hi,

I trained my model with custom Pascal VOC dataset.
This is the result.
error.PNG
How can I fix this ?

Thanks

Hi,

Could you share more information with us?

Which sample do you use?
How do you modify the config for the custom model?

Thanks.

This is config. And I am using eepstream_python_apps/apps/deepstream-test1-usbcam/deepstream_test_1_usb.py

################################################################################
[property]
gpu-id=0
net-scale-factor=0.0039215697906911373
onnx-file=../../../../samples/models/Helmet/ssd-mobilenet.onnx
model-engine-file=../../../../samples/models/Helmet/ssd-mobilenet.onnx_b1_gpu0_fp32.engine
labelfile-path=../../../../samples/models/Helmet/labels.txt
force-implicit-batch-dim=1
batch-size=1
network-mode=1
num-detected-classes=7
filter-out-class-ids=3;4;5;6
interval=0
gie-unique-id=1
output-blob-names=boxes;scores
parse-bbox-func-name=NvDsInferParseCustomONNX
custom-lib-path=/opt/nvidia/deepstream/deepstream-5.1/sources/libs/nvdsinfer_customparser/libnvds_infercustomparser.so
maintain-aspect-ratio=1

[class-attrs-all]
threshold=0.5
pre-cluster-threshold=0.7
eps=1
group-threshold=1

Hi,

Do you use a USB camera?

Could you check if this issue can be reproduced with a video?
If yes, please share the video with us for debugging.

Thanks.

resim_2021-11-09_105230

python3 deepstream_test_1.py ../../../../samples/streams/sample_720p.h264 

Onnx file and labels if needed…
ssd-mobilenet.onnx (28.6 MB)
labels.txt (57 Bytes)

Hi,

Just want to confirm the issue first.
The problem is there are too many bounding boxes. Is that correct?

Since we saw some unusual green and red regions in the first output.
But the issue is gone in the second output you shared.

Thanks.

Hi,

Problem is too many bounding boxes.

In first picture there is too many of them they just paint the screen. Its also occurs in video time to time I just couldn’t capture it.

Thanks

Hi,

We try to reproduce this issue in our environment but get some error from the post-processing.
It seems that you have a custom nvdsinfer_customparser, could you also share the source with us?

Thanks.

Hi,

I assume you are asking for this.

nvdsinfer_customclassifierparser.cpp (4.3 KB)

Thanks.

any progress on this?
could you share the steps how to train on pascal VOC?
thanks!

Hi,

No progress yet.

You can see my related question here about training.

I used cvat for annotation and export as Pascal VOC dataset for training.

Thanks.

Hi,

Sorry for the late.

In general, you should be able to run the model with Deepstream if it can be successfully deployed.
But since model has lots of variance, you may need to update some parameter or parser.

We are checking the details internally.
Will get back to you later.

Thanks.

Hi

No problem.

I also thought it might be caused by parser but I was using same parser with secondary. It was not accurate because of the data but not experienced any problems except that.

I will be waiting.

Thanks.

Hi,

We can reproduce this issue internally.

The reason of this issue is that there is no NMS (non maximum suppression) in your parser.
It is a common requirement for the detector. Please find a introduction below:

We have an example for NMS-enabled SSD parser. It should give you some idea:

$ cd /opt/nvidia/deepstream/deepstream-6.0/sources/objectDetector_SSD/
$ sudo wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/ssd-jetson_inference/ssd-jetson_inference.patch
$ sudo git apply ssd-jetson_inference.patch
$ sudo CUDA_VER=10.2 make -C nvdsinfer_custom_impl_ssd/

Thanks.