This repository is one of my projects during my three-month-internship in K2Data. I am the main contributor in a 3-people team, and the achievement has been applied in over 2,000 factories. It contains only codes because the data files are over the maximum limitation. If you want to get all the data files, or if you have any questions, please email me.
It shows how to deploy Object Detection API - Tensorflow and give specific tutorials. It is based on Object Detection. A specific example is to detect helmet, it could also explore to any other object detection tasks.
Step 1. Label - Generate the XML files based on original images, which means you locate the object in images and record the information into XML files. Label the image using lablelImg. The detailed explanation can be found here. The Training data needs to follow the organization:
--data/train
--Annotations
--JPEGImages
If you have already have the datasets containing XML and JPG files, then you can skip to Step 2.
Step 2. Generate TF-Records. TensorFlow object detection API doesn’t take XML or Image files as an input, but it needs record files to train the model. In the example, I used /data/cap_train.record and /data/cap_val.record as input. You can generate the files using /codes/create_helmet_record.py, just remember to change the dir:
# data_dir = FLAGS.data_dir
data_dir = '../data/train/'
# label_map_dict = label_map_util.get_label_map_dict(FLAGS.label_map_path)
label_map_dict = label_map_util.get_label_map_dict('../models/faster_RCNN_Inception_v2/helmet_label_map.pbtxt')
/models/faster_RCNN_Inception_resnet_v2/helmet_label_map.pbtxt is a file for labeling. Give label name i.e helmet in my example. If in case you have multiple labels, increase id number starting from 1 and give appropriate label name.
item{
id: 1
name: 'helmet'
}
num_train = int(0.8 * num_examples)
train_examples = examples_list[:num_train]
val_examples = examples_list[num_train:]
Above codes means how many datasets you split as train data and test data. I used 0.8 here.
Once the records files are ready, it's ready to train the model.
1.Choose the pre-trained model to be used. You could download Tensorflow detection model zoo, which contains various pre-trained models. My example uses /models/faster_RCNN_Inception_resnet_v2/faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28 for a better accuracy while with slower speed, and you could exchange to anyother models.
2.Remember to use the consistent config file for the same model. My example uses /models/faster_RCNN_Inception_resnet_v2/faster_rcnn_inception_resnet_v2_atrous_coco_alldata.config. Remember to change file dir in config file. Remember to change the number of classes in the file according to specific requirement. Configre Document, TensorFlow Object Detection API tutorial shows more details on how to set parameters.
3.In /codes/ folder, execute bash train.sh, remember to change the dir. You can also execute the command in Terminal. The training process requires tons of time. If you use GPU, it will occupy all of the GPU and you could use the command CUDA_VISIBLE_DEVICES=0 bash train.sh to assign a specific GPU.
# Train model
python3 ../tensorflow/models/research/train.py \
--logtostderr \
--train_dir=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802 \
--pipeline_config_path=../models/faster_RCNN_Inception_resnet_v2/faster_rcnn_inception_resnet_v2_atrous_coco_alldata.config
Before Evaluating, models must be saved. The codes is in /codes/train.sh.
# save Model
python ../tensorflow/models/research/object_detection/export_inference_graph.py \
--input_type=image_tensor \
--pipeline_config_path=../models/faster_RCNN_Inception_resnet_v2/faster_rcnn_inception_resnet_v2_atrous_coco_alldata.config \
--trained_checkpoint_prefix=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802/model.ckpt \
--output_directory=../savedModels/savedModel_100000_train_faster_RCNN_Inception_ResNet_v2_0802
The final step is to evaluate the trained model saved in ../output/ directory. There are two ways to evaluate, using eval.py to calculating the total accuracy, or using object_detection_pic.py to evaluate a single image. The following shows both two ways.
Use eval.py file and evaluate using following command:
python ../tensorflow/models/research/object_detection/eval.py \
--logtostderr \
--pipeline_config_path=../models/faster_RCNN_Inception_resnet_v2/faster_rcnn_inception_resnet_v2_atrous_coco_alldata.config \
--checkpoint_dir=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802 \
--eval_dir=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802/eval/
This will save the eval results in eval/ directory. There is a excellent tools: TensorBoard, which could visualize both training and evaling process. To launch TensorBoard (while you visualize the eval results, it would also visualize the training results at the same time), using following command in a new Terminal and locate the terminal in the file:
#To visualize the eval results
tensorboard --logdir=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802/eval/
#TO visualize the training results
tensorboard --logdir=../output/output_100000_train_faster_RCNN_Inception_ResNet_v2_0802
Instructions for TensorBoard for the beginners: TensorBoard.
The codes are in /codes/train.sh, remember to change the dir.
# Evaluate Model
python ../codes/object_detection_pic_0.75_train-test_faster_RCNN_Inception_ResNet_v2.py