ayankumarbhunia / deep-one-shot-logo-retrieval

A Deep One-Shot Network for Query-based Logo Retrieval [Pattern Recognition 2019, Elsevier]

A Deep One-Shot Network for Query-based Logo Retrieval

Ayan Kumar Bhunia, Ankan Kumar Bhunia, Shuvozit Ghose, Abhirup Das, Partha Pratim Roy, Umapada Pal “A Deep One-Shot Network for Query-based Logo Retrieval”. Accepted in Pattern Recognition, Elsevier.

We use Sigmoid layer on the output (not softmax -- this is a typo in the paper) --- by Authors

Abstract

Logo detection in real-world scene images is an important problem with applications in advertisement and marketing. Existing general-purpose object detection methods require large training data with annotations for every logo class. These methods do not satisfy the incremental demand of logo classes necessary for practical deployment since it is practically impossible to have such annotated data for new unseen logo. In this work, we develop an easy-to-implement query-based logo detection and localization system by employing a one-shot learning technique using off the shelf neural network components. Given an image of a query logo, our model searches for it within a given target image and predicts the possible location of the logo by estimating a binary segmentation mask. The proposed model consists of a conditional branch and a segmentation branch. The former gives a conditional latent representation of the given query logo which is combined with feature maps of the segmentation branch at multiple scales in order to obtain the matching location of the query logo in a target image, should it be present. Feature matching between the latent query representation and multi-scale feature maps of segmentation branch using simple concatenation operation followed by $1\times1$ convolution layer makes our model scale-invariant. Despite its simplicity, our query-based logo retrieval framework achieved superior performance in FlickrLogos-32 and TopLogos-10 dataset over different existing baselines.

Architecture

Qualitative Results

Citation

If you find this article useful in your research, please consider citing:

@article{bhunia2019deep,
  title={A Deep One-Shot Network for Query-based Logo Retrieval},
  author={Bhunia, Ayan Kumar and Bhunia, Ankan Kumar and Ghose, Shuvozit and Das, Abhirup and Roy, Partha Pratim and Pal, Umapada},
  journal={Pattern Recognition},
  pages={106965},
  year={2019},
  publisher={Elsevier}
}

deep-one-shot-logo-retrieval's Issues

Process of model training

HI,
I came across your paper and I was wondering how to set this up. Based on the setup I assumed that I can use some examples and a query Image and then the logo will be detected.
Or do I need to use the Flickr-32 dataset and train a model. Then I can use this model in order to detect the logos. Is this the way to go?

Sorry if this is a stupid question..

Weights availability

Hey Ayan! Very nice job!

Could you upload the weights from the published version? Are you planning to release some training code anytime soon?

Thanks!

Why softmax for binary cross entropy loss

1. As it is mentioned in the paper, how relevant it is to use soft max for binary cross entropy loss.? Kindly explain.
2. tf.keras.losses.bianry_crossentropy averages the loss only in the W(Width) direction. In the paper, it's mentioned to take average in H(Height) direction also. Please clarify if my understanding is correct or not.
3. Having softmax as the final layer does not show any good results. With sigmoid as the final layer the validation accuracy is reaching 98% with the validation loss of 3%. Please clarify what layer have been used in your experiments.
4. For training I have generated 154560 triplets and 90% of that has been used for training and remaining for validation. Please provide accuracy and loss statistics of your experiments.

pretrained model

could you please provide us your pre-trained weights for flickrlogo-32 dataset?

Getting false logo detection

I trained the model architecture and got good result on train and val dataset. But when I give an unknown brand logo query and its target image its detects the logo but when I give other brands query image keeping target image same even then it detects the logo. E.g I gave bata logo query image and then bata target image it detects the bata logo but when I changed the query image to samsung it still detects bata logo in target image . Where as Ideally it should give blank mask.Please help. I am stuck.

A newbie,

How to run ?

Pre-trained weights

Hey!
could you please provide us your pre-trained weights for flickrlogo-32 dataset?

Training code

Are you able to add the training code so we can see relevant hyperparameters for getting results close to the ones reported in your paper?

Thanks!