Different class label between tlt-infer and generated TensorRT engine on DeepStream

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Dear,
I’m using Nvidia TLT on docker to train my own YoloV3 model for 5 classes: bike, car, truck, minibus and bus. After training, I tested unpruned model with tlt-infer and it give correct result. I use tlt-export to generate .etlt file, and use it in Deepstream but the car class is recognize as truck class. I attach the result of tlt-infer and the result of Deepstream on .etlt file, as well as training config and Deepstream config file that i’m using. I also use tlt-converter on tlt docker notebook to generate .engine file and use in Deepstream but no luck.

yolo_train_resnet18_kitti.txt (2.0 KB)
dstest3_pgie_config.txt (1.5 KB)

deepstream
deepstream1147×888 541 KB
tltinfer
tltinfer2048×1536 1.2 MB

3

Can you share the label.txt too?

4

here is my labels.txt
labels.txt (27 Bytes)

5

Please change

bike
car
truck
minibus
bus

to

bike
bus
car
minibus
truck

and retry.

1 Like
6

hi, it work. Does it sorted by alphabet? I don’t see this info in document, or maybe I miss.

7

Yes, sorted by alphabet. In tlt user guide, frcnn part mentioned that. But yolo_v3 part is missing it.

8

thank you very much for your support, it tooks me about 1 day to find what wrong, I think we should update this info in document, because it hard to know.

9

Sorry for inconvenient. I already sync with internal team previously.
See Question on label file of deepstream_tlt_apps
Will fix it in next release.