This Tutorial is base on the link that's for windows OS I changed some commands and codes for Linux
Install Anaconda (Python 3) from link
Follow the link
conda create --name tensorflow_gpu
conda activate tensorflow_gpu
Run the following commands:
conda install pillow
conda install lxml
conda install cython
conda install contextlib2
conda install jupyter
conda install matplotlib
conda install pandas
conda install opencv
conda install tensorflow-gpu
Type python to open a python shell (Make sure you are in virtual environment)
import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))
if you see the b'Hello, TensorFlow!' output, you are good to go to the next step
Clone link Rename the folder to models
Go to models/research/object_detection and create images and inference_graph and training folders
Clone this repository and paste files in models/research/object_detection
Go to link and download your desire tensorflow zoo models.
Help: At COCO-trained models, you can choose your model that has the more COCO mAP[^1] needs more time and higher GPU but higher performance.
Extract it in models/research/object_detection folder
Goto models/research and run the command to set the python path (make sure to run the command in research dir)
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
Download protoc from GitHub
Extract the file and copy protoc file in bin folder to your models/research directory and run the command
protoc object_detection/protos/*.proto --python_out=.
In models/research directory run the command
python setup.py build
python setup.py install
In images folder models/research/object_detection/images create two directories name train and test
Copy 20% all images in the test and 80% in train directory
Notes: make sure your images are not too large it may kill the training (Use resizer.py and sizeChecker.py to resize and check your image size)
Go to this GitHub page and download Labelimg for labeling images
Install guid: clone repository and run the following command in repository directory
Run the following command to run Labelimg
python labelImg.py
Set directory for train and test and start labeling images
Run the command in models/research/object_detection/ directory to convert your XML to CSV
python xml_to_csv.py
In models/research/object_detection/ directory open generate_tfrecord.py with a text editor and set your label in class_text_to_int function then run the command to create labels file
python generate_tfrecord.py --csv_input=images/training --image_dir=image/train --out_path=train.record
python generate_tfrecord.py --csv_input=images/testing --image_dir=image/test --out_path=test.record
Create labelmap.pbtxt in models/research/object_detection/training directory and write your label for example:
item {
id: 1
name: 'your-item-1'
}
item {
id: 2
name: 'your-item-2'
}
item {
id: 3
name: 'your-item-3'
}
item {
id: 4
name: 'your-item-4'
}
item {
id: 5
name: 'your-item-5'
}
item {
id: 6
name: 'your-item-6'
}
Go to models/research/object_detection/samples/config and copy your tensorflow zoo model and paste it in models/research/object_detection/training folder
Set num_classes to your class numbers
Set fine_tune_checkpoint path to model.ckpt in your tensorflow zoo model directory
In train_input_reader block set input_path to point to train.record and label_map_path to labelmap.pbtxt
Set num_examples to number of images in testing directory
In eval_input_reader block set input_path to point at test.record and label_map_path point to labelmap.pbtxt
Goto legacy folder and run the command to start training
python train.py --logtostderr --train_dir=../training/ --pipeline_config_path=../training/<your-tensorflow-zoo model>.config
The train starts at a high loss and it will become low after training. The lower loss the better image detector.
Notes: tensorflow will save training after 5 minutes when you taught training is enough, press ctrl + c to end training
After the training procsees starts you can see the process by running the command bellow (in object_detection folder)
It shows amount of loss in training duration.
tensorboard --logdir=training/
Go to models/research/object_detection/training directory and find the highest number of model.ckpt-<highest-number>
When the train complete (loss amount is nearly 0) goto object_detection folder and run the command
Replace <higest-number> with number you found in models/research/object_detection/training directory (mine 177515)
python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/<your-tensorflow-zoo model>.config --trained_checkpoint_prefix training/model.ckpt-<higest-number> --output_directory inference_graph
Now open Object_detection_image.py with text editor and replace NUM_CLASSES with your classes number
Then create folder name test_img in object_detection and copy images you want to test your Object Detection
Run the following command
python Object_detection_image.py
press q to exit image and open next image
Open Object_detection_video.py and set NUM_CLASSES
create folder name test_video directory in object_detection and copy videos you want to test your Object Detection
Run the following command
python Object_detection_video.py
press q to exit video and open next video
Open Object_detection_webcam.py and set NUM_CLASSES
Run the following command
python Object_detection_webcam.py
press q to exit
© MRVAFA Licensed Under Apache License Version 2.0