Skip to content
deepstream 4.x samples to deploy TLT training models
C++ C Makefile
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
models Update models and README. Sep 23, 2019
nvdsinfer_customparser_frcnn_uff Update readme for oss build Nov 21, 2019
nvdsinfer_customparser_mrcnn_uff Update models and README. Sep 23, 2019
nvdsinfer_customparser_ssd_uff Update models and README. Sep 23, 2019 Init frcnn, ssd, mrcnn samples Sep 11, 2019
Makefile Update readme for oss build Nov 21, 2019
deepstream_custom.c Update models and README. Sep 23, 2019
pgie_ssd_uff_config.txt Update models and README. Sep 23, 2019

FasterRCNN, SSD, and MaskRCNN samples with Deepstream SDK

This repository provides 3 DeepStream sample apps based on NVIDIA DeepStream SDK.

These DeepStream samples support both NVIDIA Tesla and Tegra platform.

The complete pipeline for these sample apps is:



  • Deepstream SDK 4.0+ You can run deepstream-test1 sample to check Deepstream installation is successful or not.

  • TensorRT 5.1 GA

  • TensorRT OSS (release/5.1 branch) This repository depends on the TensorRT OSS plugins. Specifically, the FasterRCNN sample depends on the cropAndResizePlugin and proposalPlugin; the MaskRCNN sample depends on the ProposalLayer_TRT, PyramidROIAlign_TRT, DetectionLayer_TRT and SpecialSlice_TRT; the SSD sample depends on the batchTilePlugin. To use these plugins for the samples here, complile a new* and replace your system*.

Please note that TensorRT OSS 5.1 branch does not support cross compilation. To compile and replace the plugin,

$ git clone -b release/5.1  && cd  TensorRT
$ git submodule update --init --recursive && export TRT_SOURCE=`pwd`
$ mkdir -p build && cd build
$ wget
$ tar xvf cmake-3.13.5.tar.gz
$ cd cmake-3.13.5/ && ./configure && make && sudo make install
$ cd ..
$ /usr/local/bin/cmake .. -DTRT_BIN_DIR=`pwd`/out
$ make nvinfer_plugin -j$(nproc)
## The* will be available in the `pwd`/out folder.  Then replace the system lib with the newly built lib.
$ sudo cp /usr/lib/aarch64-linux-gnu/    /usr/lib/aarch64-linux-gnu/
$ sudo cp `pwd`/out/    /usr/lib/aarch64-linux-gnu/


  • $ export DS_SRC_PATH="Your deepstream sdk source path".
  • $ cd nvdsinfer_customparser_frcnn_uff or nvdsinfer_customparser_ssd_uff or nvdsinfer_customparser_mrcnn_uff
  • $ make
  • $ cd ..
  • $ make


We need to do some configurations before we can run these sample apps. The configuration includes two parts. One is the label file for the DNN model and the other is the DeepStream configuration file.

Label file

The label provides the list of class names for a specific DNN model trained in one of the methods mentioned above. The label varies for different apps. Details given below.

  • FasterRCNN For FasterRCNN, the label file is nvdsinfer_customparser_frcnn_uff/frcnn_labels.txt. When training the FasterRCNN model you should have an experiment specification file. The labels can be found there. For example, suppose the class_mapping field in experiment specification file looks like
class_mapping {
key: 'Car'
value: 0
class_mapping {
key: 'Van'
value: 0
class_mapping {
key: "Pedestrian"
value: 1
class_mapping {
key: "Person_sitting"
value: 1
class_mapping {
key: 'Cyclist'
value: 2
class_mapping {
key: "background"
value: 3
class_mapping {
key: "DontCare"
value: -1
class_mapping {
key: "Truck"
value: -1
class_mapping {
key: "Misc"
value: -1
class_mapping {
key: "Tram"
value: -1

We choose an arbitrary key for each number if there are more than one key that maps to the same number. And we only include the keys that map to non-negative numbers since classes mapped to negative numbers are don't-care classes. Thus, the corresponding label file would be(has to be in the same order of the numbers in the class mapping):

  • SSD The order in which the classes are listed here must match the order in which the model predicts the output. This order is derived from the order in which the objects are instantiated in the dataset_config field of the SSD experiment config file as mentioned in Transfer Learning Toolkit user guide. For example, if the dataset_config is like this:
dataset_config {
  data_sources {
    tfrecords_path: "/home/projects2_metropolis/datasets/maglev_tfrecords/ivalarge_tfrecord_qres/*"
    image_directory_path: "/home/IVAData2/datasets/ivalarge_cyclops-b"
  data_sources {
    tfrecords_path: "/home/projects2_metropolis/datasets/maglev_tfrecords/its_datasets_qres/aicities_highway/*"
    image_directory_path: "/home/projects2_metropolis/exports/IVA-0010-01_181016"
  validation_fold: 0
  image_extension: "jpg"
  target_class_mapping {
    key: "AutoMobile"
    value: "car"
  target_class_mapping {
    key: "Automobile"
    value: "car"
  target_class_mapping {
    key: "Bicycle"
    value: "bicycle"
  target_class_mapping {
    key: "Heavy Truck"
    value: "car"
  target_class_mapping {
    key: "Motorcycle"
    value: "bicycle"
  target_class_mapping {
    key: "Person"
    value: "person"


  target_class_mapping {
    key: "traffic_light"
    value: "road_sign"
  target_class_mapping {
    key: "twowheeler"
    value: "bicycle"
  target_class_mapping {
    key: "vehicle"
    value: "car"

The corresponding label file will be

  • MaskRCNN TBD

DeepStream configuration file

The DeepStream configuration file provides some parameters for DeepStream at runtime. For example, the model path, the label file path, the precision to run at for TensorRT backend, input and output node names, input dimensions, etc. For different apps, although most of the fields in the configuration file are similar, there are some minor differences. So we describe them one by one below. Please refer to DeepStream Development Guide for detailed explanations of those parameters.

Once you finish training a model with Transfer Learning Toolkit, you can run tlt-export command to generate an .etlt model. This model can be deployed on DeepStream for fast inference. The DeepStream sample app can also accept the TensorRT engine(plan) file generated by running the tlt-converter tool on the .etlt model. The TensorRT engine file is hardware dependent, while the .etlt model is not. You may specify either a TensorRT engine file or a .etlt model in the config file, as below.

  • FasterRCNN The FasterRCNN configuration file is pgie_frcnn_uff_config.txt. You may need some customization when you train your own model with TLT. A sample FasterRCNN configuration file looks like below. Each field is self-explanatory.

    A sample .etlt model is available at models/frcnn/faster_rcnn.etlt. The pb model under models/frcnn should not be used for this sample.

    # Provide the .etlt model exported by TLT or a TensorRT engine created by tlt-converter
    # If use .etlt model, please also specify the key('nvidia_tlt')
    # model-engine-file=./faster_rcnn.uff_b1_fp32.engine
    ## 0=FP32, 1=INT8, 2=FP16 mode
  • SSD The SSD configuration file is pgie_ssd_uff_config.txt.

  • MaskRCNN The MaskRCNN configuration file is pgie_mrcnn_uff_config.txt.

Run the sample app

Make sure "deepstream-test1" sample can run before running this app. Once we have built the app and finished the configuration, we can run the app, using the command mentioned below.

./deepstream-custom <config_file> <H264_file>

Known issues and Notes

  • To run FasterRCNN/SSD in fp16 mode, please replace "/opt/nvidia/deepstream/deepstream-4.0/lib/" in your platform by fp16_fix/ or fp16_fix/

  • For SSD, don't forget to set your own keep_count, keep_top_k in nvdsinfer_custombboxparser_ssd_uff.cpp for the NMS layer, if you change them in the training stage in TLT.

  • For FasterRCNN, don't forget to set your own parameters in nvdsinfer_customparser_frcnn_uff/nvdsinfer_customparser_frcnn_uff.cpp if you change them in the training stage in TLT.

  • For MaskRCNN, app can show bbox but cannot show mask in present. User can dump mask in the buffer out_mask in nvdsinfer_customparser_mrcnn_uff/nvdsinfer_custombboxparser_mrcnn_uff.cpp.

  • In function 'attach_metadata_detector()' in deepstream source code:

  1. frame scale_ratio_x/scale_ratio_y is (network width/height) / (streammux width/height)
  2. Some objects will be filtered because its width/height/top/left is beyond the source size (streammux is as source)
You can’t perform that action at this time.