Skip to content

Latest commit

 

History

History
169 lines (137 loc) · 7.74 KB

README.md

File metadata and controls

169 lines (137 loc) · 7.74 KB
Raw

VCCA for Grocery Store Dataset

This is the official implementation of our paper Using Variational Multi-view Learning for Classification of Grocery Items that was published in the Cell Press journal Patterns. This journal paper is an an extension of our WACV 2019 paper. The repository includes implementation of Variational Canonical Correlation Analysis (VCCA) for grocery item classification with the Grocery Store dataset. VCCA can make use of the web-scraped information in the dataset (i.e. iconic images and text descriptions) to learn better representations of the grocery items.

Usage

Follow the instructions below to perform experiments with the implemented models. Note that the code is written in Tensorflow 1!

See REPRODUCE for instructions to reproducing the results in the paper.

Install conda environment

Install the conda environment by executing the following command in a terminal:

conda env create -f environment.yml
conda activate vcca_grocerystore

Data

Download the Grocery Store dataset and store it wherever appropriate for you.

Github link to the Grocery Store dataset

Preprocessing of data

Before training, we have to preprocess the data. The preprocessing includes fetching image paths and labels, and creating a vocabulary for the text descriptions. The default text description length is 36 words, but this can be changed by

python ./data/preprocess_data.py --data_dir /path/to/GroceryStoreDataset/dataset \
	--save_dir /path/to/save/data

Off-the-shelf features for the natural images

We have extracted off-the-shelf features from a pre-trained DenseNet169 network. The features should be stored in the same directory as where the preprocessed data is located.

Google Drive link to download DenseNet features

Training

Selecting a model

Start by selecting which model you wish to train:

  • VCCA: vcca_xi, vcca_xiy, vcca_xw, vcca_xwy, vcca_xiw, vcca_xiwy, vcca_xy
  • VCCA-private: vcca_private_xi, vcca_private_xiy, vcca_private_xw, vcca_private_xwy
  • VAE: vae
  • SplitAutoencoder: splitae_xi, splitae_xiy, splitae_xw, splitae_xwy, splitae_xiw, splitae_xiwy, splitae_xy
  • Autoencoder: ae

The subscript indicates which data views from the dataset that are used during training:

  • x: Image features extracted from pre-trained DenseNet (download features above)
  • i: Iconic images of grocery items
  • w: Text descriptions of grocery items
  • y: Class labels of the natural images

The VAE and Autoencoder model names do not use a subscript because they only use the image features x. If selecting a model without y, then classification is performed by training a softmax classifier on the latent representations of the model.

Scaling weights for reconstruction losses

We can choose to scale the reconstruction losses for each view by passing a number to the following arguments to train.py:

  • --lambda_x: Scaling weight for image feature loss
  • --lambda_i: Scaling weight for iconic image loss
  • --lambda_i: Scaling weight for text descriptions loss
  • --lambda_y: Scaling weight for class labels

The default values for all scaling weights is 1.

Training the selected model

As an example, we train the model vcca_xiwy by executing:

python train.py --data_path /path/to/processed_data --model_name vcca_xiwy

If we would like to change the scaling weights of the model, we pass value of the scaling weights as arguments by executing:

python train.py --data_path /path/to/processed_data --model_name vcca_xiwy --lambda_i 1000 --lambda_w 1000 --lambda_y 1000

The file clf_metrics.txt includes the fine- and coarse-grained accuracies predicted by the used classifier.

For saving the trained model, pass the argument --save_model 1. You can also specify the directory where the model should be saved with the argument --model_dir /path/to/saved_model. If the softmax classifier is used, then it is stored in /path/to/saved_model/saved_classifier.

Test

You can load a trained model in the script test.py to

  • compute the fine- and coarse-grained accuracy
  • plot the latent representations in 2D
  • decode iconic images from natural images and compute metrics for the decoded images (if iconic image decoder was used in model)

Run the script by executing:

python test.py --data_path /path/to/processed_data --model_dir /path/to/saved_model --model_name MODEL_NAME

The metrics and images are saved in the directory saved_images_and_metrics by default. The directory can be passed as argument with --save_dir /path/to/new_name_for_saved_metrics_and_images

Latent representation in 2D using PCA

Fig 1: Latent representations plotted with the corresponding iconic images of models vae and vcca_xiwy in the left and right figure respectively.

Decoded iconic images

Fig. 2: Four examples of decoded iconic images from model vcca_xiwy by encoding the natural image and decoding the retrieved latent representation through the iconic image decoder. The images are structured in the following order: 1) natural image, 2) true iconic image, 3) decoded iconic image.

Citation

If you use this code or the Grocery Store dataset for your research, please cite our papers:

@article{klasson2020using,
  title={Using Variational Multi-view Learning for Classification of Grocery Items},
  author={Klasson, Marcus and Zhang, Cheng and Kjellstr{\"o}m, Hedvig},
  journal={Patterns},
  volume={1},
  number={8},
  pages={100143},
  year={2020},
  publisher={Elsevier}
}

@inproceedings{klasson2019hierarchical,
  title={A Hierarchical Grocery Store Image Dataset with Visual and Semantic Labels},
  author={Klasson, Marcus and Zhang, Cheng and Kjellstr{\"o}m, Hedvig},
  booktitle={IEEE Winter Conference on Applications of Computer Vision (WACV)},
  year={2019}
}

Acknowledgement

This research was funded by Stiftelsen Promobilia in Stockholm, Sweden.