Skip to content

Build a computer vision-based technology to process and detect the potholes present in an image.

License

Notifications You must be signed in to change notification settings

hariprasath-v/AV-Dataverse-Hack-Build-an-AI-Model-to-Save-Lives

Repository files navigation

AV-Dataverse-Hack-Build-an-AI-Model-to-Save-Lives

Competition hosted on Analyticsvidhya

About

Build a computer vision-based technology to process and detect the potholes present in an image.

This is my first object detection - computer vision hackathon.

From the competition starter notebook, I have tried the same PyTorch - fasterrcnn resnet50 fpn object detection model, and without the proper image processing or data augmentation the model not performed well and didn't learn any signals.

Final Competition score is 0.0060050745

Leaderboard Rank is 72/77

Evaluation Metric is mAP@[.5,.95]

Post-competition I personally tried the yolo object detection model with wandb logging. The model scored 38%(100*mAP[.5,.95]) on test dataset.

File information

  • av-dataverse-hack-build-an-ai-model-to-save-lives- EDA.ipynb Open in Kaggle

    Basic Exploratory Data Analysis

    Packages Used,

     * seaborn
     * Pandas
     * Numpy
     * Matplotlib
     * PIL
     * cv2
     * os
     * distance
     * imagehash
     * time
     * itertools
    

    Extract basic information about the images(width, height, color mode) and analyzed the information through visualization in the following methods.

    Total pothole wise image samples.

    Alt text

    RGB color distribution analysis.

    Alt text

    Find the similar images using different hashing algorithms.

      * Average hashing - Total matched images 284
      * Perceptual hashing - Total matched images 80
      * Difference hashing - Total matched images 80
      * Wavelet hashing - Total matched images 280
      * Color hashing - Total matched images 217120
    

    From the above various image hashing algorithm, the perceptual hashing, and difference hashing algorithms significantly find similar images based on the hash value.

    Alt text