/Object-Detection-with-YOLO

Primary LanguageJupyter Notebook

Object Detection with YOLO

  • Data Preparation: Given approximately 167 Koala Images and instructions to label those images and build a model on YOLO. I labeled images with the open-source online annotator LabelMe. Create a bounding box for every image and have 5 values one each such as label, x_min, y_min, x_max, y_max. After annotating all the images upload on Google Colab for further investigation. Import every JSON file and append those together to create a Txt file for each. Each txt file will contain 5 values and we know, but we need different values like x_mid, y_mid, width, and height. Calculate and process that Txt files, and finally create Txt files with expected values. Now match every image with a label to find the right image for processed images. Having 163 labels from 167 images discard 4 images and map those images with corresponding images. Now divide images with train and test and on train images use fold cross-validation with n_splits=3. Create a directory and save processed images, labels, and classes.

  • YoLo Format Creation: To train the YOLO model, create a YAML file that contains the train, valid, number of classes, and classes. For validation 1 folds and the rest of the two folds are used as training data.

  • Import YOLO: You only look Once is the full form of YOLO. For this dataset, I used the yolov5 model with small weights. Clone the model with the current directory and install the requirements.txt. After checking the installation and running the pre-define detect the image. Now time to train the model, FYI I trained the model three times using different fold and image sizes. Universally I used batch size=16, epochs=50, images_size=220 one time and two times 640. Not done any Hyperparameter tuning which is a necessary point to train any pre-train model. If did that then the box loss would be decreased besides the evaluation metrics would be increased. But on top of that got better results on difference fold.

Fold box_Loss Precision Recall mAP
1 0.0209 0.963 0.978 0.993
1 0.02092 0.982 0.993 0.995
3 0.0199 0.992 0.993 0.995
  • Create An API: I created an API for better usability of the model with FastAPI. where set an endpoint to predict the bounding box of the image which I want to select.

END NOTEI learned a lot of things while doing this task, The preprocessing steps were interesting, and also selected a model for small images, We all know Yolo works better on high-resolution images like 640x640 or 1280x1280. Find out the perfect model and parameter is more interesting part while training a model like batch_size, optimizer, which weights better, and so on. Faced different types of challenges while creating API Locally with FastAPI finally solved those issues.