/Object-Detection-SSD-step-by-step

Step-by-step SSD training on own dataset

Object-Detection-SSD-step-by-step

Step-by-step SSD training on own dataset

Dependencies Installation

  1. Clone Tensorflow Object Detection API Repo
git clone https://github.com/tensorflow/models.git
  1. Install dependencies required in Requirement

Data Preparation

  1. New data can be labeled using lableImg Repo
    Note: label files should be in VOC format (.xml files)
  2. Split data into train and validation sets:

Split randomly with 80% training and 20% validation Convert .xml files into train.tfrecord and valid.tfrecord files

Train SSD with own dataset

  1. Modify .config file in /samples/configs/<>.config
  2. Train SSD as show in Tensorflow Repo
# From the tensorflow/models/research/ directory
PIPELINE_CONFIG_PATH={path to pipeline config file}
MODEL_DIR={path to model directory}
NUM_TRAIN_STEPS=50000
SAMPLE_1_OF_N_EVAL_EXAMPLES=1
python object_detection/model_main.py \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --model_dir=${MODEL_DIR} \
    --num_train_steps=${NUM_TRAIN_STEPS} \
    --sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \
    --alsologtostderr

Inference

  1. Turn trained model into frozen graph
  2. Convert model into TensorRT version for inference

Convert model into TensorRT

  1. Hardware: Jetson Nano
  2. Installation:

Setup Jetson Nano as required in NVIDIA Site Build tensorflow 1.12.2 sited in jkjung-avt.github

sudo pip3 uninstall -y tensorboard tensorflow
cd <PROJECT_DIR>
git clone https://github.com/jkjung-avt/jetson_nano.git
cd jetson_nano
./install_protobuf-3.6.1.sh
./install_bazel-0.15.2.sh
./install_tensorflow-1.12.2.sh

Note when building tensorflow on Jetson Nano: error: not found line #define #define NV_TENSORRT_SONAME_MAJOR in /usr/include/aar.../Nvinfer.h can be solve by placing the following lines in Nvinfer.h file.

#define 	NV_TENSORRT_SONAME_MAJOR   4 
#define 	NV_TENSORRT_SONAME_MINOR   0
#define 	NV_TENSORRT_SONAME_PATCH   1