How to detect small object with deepstream?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Currently, we found that the demo application can detect object, but it needs quite a lot of pixels for detection. Is there any way to detect small object with less pixeles, such as 16x16 pixels?

3

What does this mean?

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU)
• DeepStream Version
• JetPack Version (valid for Jetson only)
• TensorRT Version
• NVIDIA GPU Driver Version (valid for GPU only)

4

The object almost needs 1/4 of the screen 1920x1080P, which is 480x270 pixels.

Jetson Orin Nano 8GB Devkit
DS 7.1
Jetpack 6.2 L4T 36.4.3
TensorRT: 10.3.0.30

5

What is your model’s input dimension(size)?

6

I’m using the default model from sample code. Is it mean I have to train my model? Or any specific small model files that I can use for test?

7

Which sample?

8

Try this one, changed from python deepstream_test2, which merged some deepstream_test3 code for performance issue.

git clone https://github.com/SnapDragonfly/jetson-fpv.git
cd jetson-fpv/utils/deepstream
ln -sf /opt/nvidia/deepstream/deepstream/samples/ samples
cd ../../
python3 ./utils/deepstream/deepstream_NvDCF.py -i rtp://@:5600
9

So you are using the TAO pretrained models in DeepStream sample. TrafficCamNet | NVIDIA NGC 's input resolution is 960x544, it is not good at detecting small objects. Please customize and train your own model to detect small objects.

10

@Fiona.Chen Well, I really don’t have much experience on AI training/inference.

Just try to use this AI edge device on fpv area, Can you give some links for small object training tutorial? And I can quick learn step by step.

11

DeepStream is a multimedia inferencing SDK which just deploy the models. If you want to use the TAO models and train the TAO models, please refer to TAO Toolkit | NVIDIA Developer. The TAO forum is Latest Intelligent Video Analytics/TAO Toolkit topics - NVIDIA Developer Forums

12

Is there any ready made mode for small vechicle/person?

13

If you want to use the TAO models and train the TAO models, please refer to TAO Toolkit | NVIDIA Developer. The TAO forum is Latest Intelligent Video Analytics/TAO Toolkit topics - NVIDIA Developer Forums

14

I don’t want to train the model myself. I hope to use some ready trained model. Is there any ready made mode for small vechicle/person?

15

No. DeepStream doesn’t provide any ready made model. Please consult in TAO forum or you need to look for some other open source models by yourself.

16

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.