No objects detected with retrained SSD model from Transfer learning toolkit

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

I trained, pruned, and retrained an SSD model from the transfer learning toolkit. I’ve updated and compiled the necessary libraries to run the *.etlt model on Deepstream-test1-app. The application runs but it is not detecting any cars. I used the sample_720p.h264 video source and I’ve verified on the tlt-streamanalytics ssd jupyter notebook that the model can detect cars the sample_720p.h264 frame. The threshold is set to 0.2. What am I doing wrong?

Thanks

2

Move this topic from DS forum to TLT forum.

3

Hi shaun,
Please paste the running command and also the config file for Deepstream-test1-app.

4

Also, please refer to https://devtalk.nvidia.com/default/topic/1070464/transfer-learning-toolkit/use-of-deepstream-4-0-2-tlt-encoded-model-to-avoid-using-tlt-converter/

5

Hello Morganh,

This is my config file with the key removed:

[property]
gpu-id=0
net-scale-factor=0.0039215697906911373
batch-size=1
network-mode=0
num-detected-classes=3
interval=0
gie-unique-id=1
process-mode=1
model-color-format=0
classifier-threshold=0.2

labelfile-path=./labels.txt
#model-engine-file=./ssd_resnet18_epoch_180_pp6.etlt_b1_fp32.engine
tlt-encoded-model=./ssd_resnet18_epoch_180_pp6.etlt
tlt-model-key=(removed)
output-blob-names=NMS
uff-input-dims=3;384;1248;0
uff-input-blob-name=Input

enable-dbscan=1
parse-bbox-func-name=NvDsInferParseCustomSSDUff
custom-lib-path=/root/deepstream_4.x_apps/nvdsinfer_customparser_ssd_uff/libnvds_infercustomparser_ssd_uff.so

[class-attrs-all]
threshold=0.3
eps=0.2
group-threshold=1

#######################################
the command I used is ./deepstream-test1-app ~/deepstream_sdk_v4.0.2_x86_64/samples/streams/sample_720p.h264.

I know that the osd_sink_pad_buffer_probe is for Vehicle, TwoWheeler, Person and Roadsign but even if the label is wrong I should still see some object detected with bounding boxes but wrong label.

Thanks for the help

6

Please check

  1. Your key is correct, without any additional [space] character in the end of line
  2. check if num-detected-classes is correct
  3. make sure the label.txt have the classes which you have trained. And the quantity of classes are correct.
7

Morganh,

The key is correct so are the num-detected-classes and label.txt. Wouldn’t the application not run if the key is incorrect? And would crash if the num-detected-classes is less than the actual number of classification the model has been trained for? I’m not seeing any application error. The application runs fine, I just do see any object detection even with the threshold set at 0.3

8

Hi shaun,
May I know if you were training KITTI dataset? Which classes did you train?
Could you please also paste the label.txt here? Thanks.

9

Hi shaun,

We haven’t heard back from you in a couple weeks, so marking this topic closed.
Please open a new forum topic when you are ready and we’ll pick it up there.