Modifying the DetectNet model on Jetson Nano

[rose.louis.mail]

Hi,

I am using the pretrained ‘ssd-mobilenet-v2’ on the Jetson Nano (4GB) to detect cars on a video.

I would like to only detect cars, and not the other classes the model is trained on, like people or boats, etc…

I saw that this is possible by retraining the model on part of the MS COCO dataset, using Transfer Learning.

But currently, I don’t have access to enough memory or time on my hands to retrain the model in such a way.

Is there a way to limit the detectnet model to detect only some object classes ? I could not find any way to do so in the documentation.

Even if the model does compute for all classes, a way to delete those unwanted detections from the detection list, before they appear on the image, would be much appreciated.

Thanks in advance,

Louis

[dusty_nv]

Hi @rose.louis.mail, you can disable certain object classes without retraining the model by setting those class names to void in the label file. In the case of ssd-mobilenet-v2, that file is found at:

jetson-inference/data/networks/SSD-Mobilenet-v2/ssd_coco_labels.txt

So you would change every line in that file to void except for the line that says car

[rose.louis.mail]

It works perfectly, thanks!