Hi, I followed the AI Hello world project with the Jetson Nano.
When I run the “Locating Objects with DetectNet” on the dataset that was provided, I got good detections.
When I tried to run it on my own pictures to recognize cars and people, I got either no detection, false detection, or partial detection.
what might be the cause of the problem and any suggestions to what could solve it?
Thank you.
Hi @giliti.gilad, are many of your images from that elevated vantage point of a security camera? The MS COCO dataset that the SSD-Mobilenet-v2 model was trained on has mostly images from a first-person view (i.e. from the ground)
You may be interested in the TLT PeopleNet model and other models from the Transfer Learning Toolkit which have been trained on data more akin to yours. You can run these models through DeepStream and you can also fine-tune it with your own data.
BTW, since you only want to find cars/people but the MS COCO model has 90 object classes, you can go into the labels.txt and replace the class names with void that you want to disable. It would not then spuriously detect ‘airplane’, for example.
can you refer me to the commands to run PeopleNet (instead of SSD-Mobilent)?
thanks
The TLT PeopleNet model runs through DeepStream, and you can find the commands to deploy it with DeepStream on this page: