Object Counting from an Image:

Counting animals from the Image using mask-rcnn and Yolo with GUI based on Django.

Tools Used:

• Framework: Django
• Language: Python
• Lybraries: opencv, keras
• Techniques: Yolo, Maskrcnn

1) Initial Setup:

• Download and Extract Zip (DjangoProject-master) * Install Mask-RCNN * Download Pretrained model weights of yolo and mask-rcnn

1.1) Install maskrcnn:

• Go to https://github.com/matterport/Mask_RCNN and download it.
• Extract the file and go the “Mask_RCNN-master” folder and open terminal in that folder.
• Run the below command in the terminal.
• Python setup.py install

1.2) Download weights:

• Download mask-rcnn Pretrained weights from here ((https://drive.google.com/file/d/1l1_N6Ue1p9rFwAUxfeZTpVpzuu0l1el-/view?usp=sharing)) And push that file to “DjangoProject/keras_mask_rcnn/”.

• Download YOLO model Pretrained weights from here ((https://drive.google.com/open?id=1gAQcJec6ExidEuGf2dCbnG9OJ6ynhXg0)) And push that file to “DjangoProject/Yolo_model/model_data/”

2) How to start the project:

• Set working directory to "DjangoProject"

• Open “Requirements.txt” files and install Libraries

• Run server by below command

  • python manage.py makemigrations
  • python manage.py migrate
  • python manage.py runserver

• Open browser(any) and enter below URL in URL-Box

  • 127.0.0.1:8000/image_upload

• Home Screen

  • Choose a file with and click on “Upload Image”
  • Now Process will starts

• Two Techniques are used (Yolo, MaskRCNN)

• “Yolo” will takes 10 seconds to detect all objects(animals) from the image where “MaskRcnn” will takes 30 to 40 seconds on My CPU system.

  • It will fast on detecting on GPU system.
  • We can decrease the prediction time also by loading the model only once.

• Default Technique is MaskRCNN

  • If you want to activate YOLO, just activate commented code and deactivate default activated code(by adding and remove comments in appropriate places).Refer below image.

To

After prediction, the results

Above screens, output has some bounding boxes on it. Those are prediction results(localization of the animals).

Count of animals = count of bounding boxes

Note: Please wait until Prediction results are out.(takes a long time based on techniques that I have used.) Special Case: For above input image, there is no bounding boxes in the output image.because this project only detects animals and gives bounding boxes to them. Rest of the objects will be ignored.

3)Improvements:

3.1) GPU systems

• For now time complexity is quite high because of I’m having CPU system. This project can run quickly in GPU systems.

3.2) Loading Trained model Only Once

• Every time model is to be loaded. Because, Django will not clears sessions automatically. We have to do some ground work on Django, then we can able to decrease time for predictions.

3.3) We can use other Multi label detectors like

• SSD, fast-rcnn, fast-rcnn for more accurate results.

3.4) Training on own dataset

• We can gather our own dataset related to all animals and train it by any one of the “Multi object detectors” like SSD, YOLO, Faster-RCNN for more accuracy.

Contact Details:

• Name: Rajashekhar.G
• Linkedin prof: https://www.linkedin.com/in/rajashekhar-gugulothu/
• Email: rajharry418@gmail.com