Deepstream python apps using TAO Faster RCNN model trained on COCO bug with class attributes


I deploy my TAO Faster RCNN model trained on 80 class COCO dataset using Deepstream with tracker. I want the tracker to only track person class. Therefore, I set the pre-cluster-threshold=1.0 for all the classes in the Deepstream_config file frccn_res50_test_pgie_config.txt except the person class which is class_number 49 as per labels2.txt (see attached files).

However, this works for all the classes except class-0. If I set the pre_cluster_threshold=1.0 for class-0 (aeroplanes), deepstream does not generate any bbox. If I comment this line out, it works fine but it also detects + tracks aeroplanes which I do not want. From my understanding, the threshold value for class-49 (person class) and class-0 (aeroplane) seem to somehow getting linked.

Please provide complete information as applicable to your setup.

frcnn_res50_test_pgie_config.txt (7.3 KB)
labels2.txt (617 Bytes)

Is there other component that need the class except person?
If there is not, you could use “filter-out-class-ids” in nvinfer plugin to filter out the objects of other classes.

Thank you, filter-out-class-ids works like a charm!

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