1 of 10

HackaTUM 2017

BlackHats

Sarthak Gupta,

Virendra Kumar Pathak,

Jiaxian Fan,

Fahad Mansoor

2 of 10

Task

  • Classify and localize logos

​

  • No of unique classes over 55

​

  • No of images 45 thousand

3 of 10

Stills

When our network likes us

And when it doesn’t

4 of 10

Approaches

  • We use Multi Layered Convolutional Neural Networks to detect whether there is a logo in the image or not.

​

  • We had a 90-10 training and validation split

​

  • And we then draw a bounding box on the logo.

5 of 10

Data Explained

Positive Examples

Negative Examples

  • We extracted images with and without logos, then processed them and then used them to train our network

Processed Images

6 of 10

Model Performance

  • Training Accuracy
  • Loss

7 of 10

Difficulties

  • Lack of computational resources and time
    • We weren’t able to train and do proper localization due to the the non availability of GPUs

​

  • Inconsistent image sizes

​

  • We were unable to train resnet152 and RCNN because of lack of computational resources

​

8 of 10

Moving Forward

  • Use Region proposal networks to classify the image and find the location of a logo anywhere in the image. (we have the architecture in place with all the required modifications to the original paper)

​

​

9 of 10

Moving Forward

  • Train Conv Nets on Logo dataset to predict if and what kind of logo is present in the prediction made by current architecture.

​

  • Try various other optimization techniques to improve prediction accuracy.

10 of 10

Questions?