Car color secondary classifier deployed in Deepstream5.1, performed low classifying efficiency,and why

Please provide complete information as applicable to your setup.

hardware tested on jetson nano and agx xvaier;


secondary classifier:car color classifier

I tried to enable car color classifier linked behind the primary object detection and tracker , and tested its classifying performance in Jetnano and xvaier respectively, unfortunately,it didn’ t performed so well as i expected, it only can classify correct color approximately mere 10% of overall detected vehicles, as i counted in the app.

all correctly classifying nearly distribute the center of images,

By the way, i retrained the primary model, it can correctly detect up tp 93% of vehicles running on the road, and another thing is the camera was mounted 6.5m height, namely, all vehicles are top views. i don’t know could this top view contribute unsafisfied classifying results. (ps, i only can refer to the vehicleMakenet and vehicletypenet’s instructions, they both mentioned Limitations as follows:

Top view

This model is design to classify detected cars from DashCamNet model with cars visible at a 5ft height, with visible front, back or side facia. The model doesn’t classify well for cars seen from top view.)

so my question:
1\ are there any limitations of car color models if deployed it in the deepstream5.1 ?
2\ how can i improve the performance of the car color model via slightly tuning the config.txt?
thanks in advance

Will check and get back to you

sorry for late!

Any news about the color classifier? i 've reduced the thresholds of color classifier from 0.51 to 0.1, then the number of color detected vehicles have increased by approximately 60%, however the accuracy was reduced as i assumed, so , could you please give me a link of new color models which can be downloaded from NGC, or somewhere?

i 've noticed that the deepstream5.1 version have excluded the color classifier, and only keeps the car maker and car type classifier there

Hi @sainttelant ,
Sorry for late response! Yes, you need to retrain the model with top view images since the model in DS was trained by different image set.

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.