Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 29 views

    Car Damage Detective: Assessing Car Damage With Convolutional Neural Networks Ting Neo

    Uploaded by

    Pravin Kanna
    This document describes research using convolutional neural networks (CNNs) to classify vehicle damage from images in order to streamline the auto insurance claims process. The researcher trained three CNN models - one to detect damaged vs undamaged vehicles, one for damage location (front, rear, side), and one for damage severity (minor, moderate, severe). The models achieved 92%, 79%, and 71% accuracy respectively, even when trained on a small dataset of only 1,150 damaged vehicle images. The researcher concludes that CNNs are effective for evaluating vehicle damage and suggests that predicting repair costs may also be possible with a higher quality dataset.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Car Damage Detective: Assessing Car Damage With Convolutional Neural Networks Ting Neo

    Uploaded by

    Pravin Kanna
    29 views11 pages
    This document describes research using convolutional neural networks (CNNs) to classify vehicle damage from images in order to streamline the auto insurance claims process. The researcher trained three CNN models - one to detect damaged vs undamaged vehicles, one for damage location (front, rear, side), and one for damage severity (minor, moderate, severe). The models achieved 92%, 79%, and 71% accuracy respectively, even when trained on a small dataset of only 1,150 damaged vehicle images. The researcher concludes that CNNs are effective for evaluating vehicle damage and suggests that predicting repair costs may also be possible with a higher quality dataset.

    Original Title

    neokt-car-damage-detective-121416

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document describes research using convolutional neural networks (CNNs) to classify vehicle damage from images in order to streamline the auto insurance claims process. The researcher trained three CNN models - one to detect damaged vs undamaged vehicles, one for damage location (front, rear, side), and one for damage severity (minor, moderate, severe). The models achieved 92%, 79%, and 71% accuracy respectively, even when trained on a small dataset of only 1,150 damaged vehicle images. The researcher concludes that CNNs are effective for evaluating vehicle damage and suggests that predicting repair costs may also be possible with a higher quality dataset.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    29 views11 pages

    Car Damage Detective: Assessing Car Damage With Convolutional Neural Networks Ting Neo

    Uploaded by

    Pravin Kanna
    This document describes research using convolutional neural networks (CNNs) to classify vehicle damage from images in order to streamline the auto insurance claims process. The researcher trained three CNN models - one to detect damaged vs undamaged vehicles, one for damage location (front, rear, side), and one for damage severity (minor, moderate, severe). The models achieved 92%, 79%, and 71% accuracy respectively, even when trained on a small dataset of only 1,150 damaged vehicle images. The researcher concludes that CNNs are effective for evaluating vehicle damage and suggests that predicting repair costs may also be possible with a higher quality dataset.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 11

    Car Damage Detective

    Assessing Car Damage with


    Convolutional Neural Networks

    Ting Neo
    Metis NYCDS9
    Dec 13, 2016
    Car accidents are stressful and the auto
    claims process is ripe for disruption

    2
    Car accidents are stressful and the auto
    claims process is ripe for disruption

    snap!

    $
    Objective: Use computer vision to
    accurately classify vehicle damage and
    facilitate claims triage
    3
    Can computer vision “see” car damage?
    Convolutional Neural Networks (CNNs)
    ⬥ VGG16 trained on Imagenet (14M images, 1000 classes)

    + Train a fully-connected classifier on top of a pre-trained CNN


    ⬥ Allows training a CNN with as little as ~300 images per class
    Transfer
    learning + Fine-tuning
    ⬥ Tweaking the last convolutional block and retraining fully
    connected classifier with a very slow learning rate
    Trained on
    bottleneck
    features

    Retrain
    Fine-tune
    Fully connected
    last group
    classifier
    VGG16 of layers

    architecture

    4
    Developing a pipeline
    Gate 1 Gate 2 Classification Results

    Model 2
    3 classes

    VGG16
    Accept results Model 1 Front
    User over a top-5 2 classes Returned
    Submits threshold Location? Rear to User
    Photo of and Third
    Damage Parties
    Side
    Yes Yes
    Car? Damage?

    Minor

    No No
    Severity? Moderate

    Not a car Undamaged Severe

    Retry Retry

    Model 3
    3 classes
    5
    Tools and frameworks used
    Python
    Programming
    Language

    Data Collection Model Development Web Development

    Selenium Amazon EC2 Amazon EC2


    Webscraper GPU Instance GPU Instance

    Flask
    Google Images Keras
    Deep Learning Python Web
    Data Source
    Library Framework

    Bootstrap
    Stanford Car Theano
    HTML, CSS,
    Image Dataset Deep Learning
    Javascript
    Data Source Library
    Framework

    1,150
    Undamaged cars

    1,150
    Damaged cars
    6
    Evaluating pipeline components
    Model 1: Model 2: Model 3:
    Damaged or whole? Damage location Damage severity
    92% accuracy 79% accuracy 71% accuracy
    91% precision 76% precision 71% precision
    90% recall 76% recall 69% recall

    Diagnosis: Diagnosis: Diagnosis:


    More damaged cars Ambiguous closeups of Subjective labeling
    identified as whole front/rear damage criteria

    Choice of precision-recall threshold depends on perspective!


    7
    Car Damage Detective: Live!
    http://bit.do/car-damage-detective

    8
    http://bit.do/car-damage-detective

    9
    Conclusions and future work

    ◈ Convolutional neural networks are accurate at evaluating


    car damage - even trained on only 1,150 damaged car
    images

    ◈ With a higher quality dataset (including make and model,


    location information, repair cost, etc.), we could go one step
    further and predict the cost of damage based on the image

    10
    Thank you!
    Assessing Car Damage with
    Convolutional Neural Networks
    Ting Neo
    http://bit.do/car-damage-detective
    neo.kaiting@gmail.com
    github.com/neokt
    linkedin.com/in/kai-ting-neo

    You might also like