EasyRFP: An Easy to Use Edge Computing Toolkit for Real-Time Field Phenotyping

We propose EasyRFP, an edge computing toolkit for real-time field phenotyping. Recent advances in deep learning have catalysed rapid progress in high throughput field phenotyping. Much research has been dedicated towards developing accurate and cost effective deep learning models to capture phenotyping traits such as plant stress, yield and plant growth stages.

However, there is a shortage of software tools to promote the usage of such intelligent methods among plant phenotyping practitioners and researchers. To bridge this gap, we developed this, a Flask backend, Angular frontend software toolkit. Broadly speaking, our toolkit can be interfaced with a commercial GPU enabled micro computer (such as NVIDIA Jetson) and a digital camera. Precisely, our toolkit can be used to capture images and extract phenotypic traits in both real-time and in scheduled mode. Currently, we support classification, detection and instance segmentation tasks.

Demonstration (Video)

Installation

Toolkit is tested only on Python 3.6. Works on both Windows and Linux systems.

To use the basic template of the toolkit (without pretrained models)

gphoto2==2.2.2
opencv-python==4.4.0
Flask==1.1.2
Flask-Cors==3.0.8
Flask-SocketIO==4.3.1
onnx==1.7.0
onnxruntime==1.5.2

NodeJS & Angular installation:

> pip install nodeenv
> nodeenv -p
> npm install

Execution

Run Flask Backend Server

Inside api/ do

Linux:

> export FLASK_APP=api
> flask run

Windows:

> set FLASK_APP=api
> flask run

Run Angular Frontend Server

Inside client-app/ do

> ng serve

Access the UI at http://localhost:4200/. If you port it to your local machine with --host=0.0.0.0 command, access UI at http://192.168.X.X:4200/. Also, please change the SOCKET_ENDPOINT variable accordingly with your IP address in client-app/src/environments/environment.ts

Some More Configuration

IMAGE:
  # Resizing is important since the DSLR captured images are huge (5000x4000 approximately) 
  # The Flask server could die trying to work with such huge images
  RESIZE_WIDTH: 640 # in pixels
  RESIZE_HEIGHT: 640 # in pixels
IO:
  LOGS_SAVE_DIR: logs/ # created inside api/
EMAIL:
  # While you are free to use any email id, you are recommended to create
  # a throwaway email id at any provider supporting SMTP (ex: Gmail).
  # The only purpose of this email id is to send emails on behalf of the Jetson.
  # Make sure you don't have any personal information connected to this email id
  # to ensure security. Please know that you have been warned.
  SRC_EMAIL_ID: [email protected]
  SRC_EMAIL_PASSWORD: dummypassword
  SMTP_SERVER: smtp.yandex.com #smtp server of your email provider -> ex: smtp.gmail.com
  SMTP_PORT: 465 # ex: 587
SESSION_NAME: your_desired_name # or auto (session folder will be timestamp string if auto) 

How To Add Your Own Task

• Create task python file in tasks directory, say dummy_task.py
• Declare a class whose constructor takes a cv2 image as input, say DummyTaskClass.
• Include a method perform_task() that uses the input image, trained model, and your own inference code to output result appropriately.
  • If it is a detection task, the method must return a list of tuples, where each tuple looks as follows: (classname, [x, y, width, height])
  • If it is a classification task, the method must return a tuple as follows: (class_name, class_probability)
  • If it is a segmentation task, the method must return a mask image (torch tensor) of size equal to the input: torch.Tensor
  • Please refer to tasks/wheat_detection.py and task/leaf_disease.py to know more.
• Once DummyTaskClass().perform_task() is configured to work as above, import the class in tasks/task_bank.py
  • Something like from .dummy_task import DummyTaskClass
  • Add a simple if statement inside the inference() method to work based on the task name (provided in YAML) and you are good to go.
  • The task_bank.py could look as follows

  from .some_other_task import SomeOtherTask
  ##### Newly added 
  from .dummy_task import DummyTaskClass
  def inference(img, task_name):
      task_obj = None
      if task_name == `SomeOtherTask`:
          task_obj = SomeOtherTask(img)
      ##### Newly added
      elif task_name == 'DummyTaskName':
          task_obj = DummyTaskClass(img)
      
      outputs = task_obj.perform_task()
      return outputs
  

• Finally, add the newly added task to tasks/alltasks.yaml as follows

  - name: DummyTaskName
    display_name: Dummy Task
    task_type: classification (or detection/segmentation)
  

Pretrained Models

• Wheat Ear Detection - FasterRCNN ResNet101 (480 MB)
• Leaf Disease Classification - ResNet18 (PyTorch) and ResNet18 (ONNX) (45 MB)
• Leaf Segmentation - MaskRCNN ResNet50 (350 MB)

Add these models to the project and provide the path in tasks/wheat_detection.py or tasks/leaf_disease.py or tasks/leaf_segmentation.py. Please check the training code for these three tasks at tasks/models/.

To use/test the provided pretrained models, you will need

detectron2==0.1.3+cu101
torch==1.5.0+cu101
torchvision==0.6.0+cu101

To install PyTorch and torchvision on the NVIDIA Jetson, you can follow the installation instructions provided here. Detectron2 can be installed with the following command:

python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

Citation

@article{easyrfp2020,
    Author = {Sai Vikas Desai, Akshay L Chandra, Masayuki Hirafuji, Seishi Ninomiya, Vineeth N Balasubramanian, Wei Guo},
    Title = {EasyRFP: An Easy to Use Edge Computing Toolkit for Real-Time Field Phenotyping},
    Journal = {https://github.com/lab1055/easy-rfp},
    Year = {2020}
}