Deepstream YOLO doesn't work

Hi,

I’m using deepstream, but yolov3-tiny on jetson nano doesn’t work :
the video is working, but I got really few bounding boxes.

Sincerely,

What about other config? For example yolov2, yolov2-tiny ?
Do you have some changes?

Hi @ChrisDing ,

After trying with differents parameters, I suppose it’s coming from my camera or maybe a problem of luminosity.
I test my camera with a window behind me.
I have a usb camera with a 720p resolution.

I tried with my curtains closed and with artificial light and it seems to be a little better, the bounding boxes seems to be more persistent. But it’s not this type of quality :


At the end the program I want to create have to work in real light, so do you think my camera is the only problem?

Sincerely,

Hi,

I tried with a new camera with better resolution but I still have a bad detection of objects, is anyone would have some advice to improve detection, parameters of deepstream/YOLOv3-tiny, or parameters for the camera?

Sincerely,

AFAIK, the first video is using YoloV3, and second video is using DetectNet_V2.
Not sure what’s the resolution of the original camera, if it’s not too low, higher resolution should not help too much since the video will be resized to the small input resolution of the network (you could find the resolution in the nvinfer configuration file).
Certainly, the camera quality will affect the detection result, maybe you could try a camera with HDR.