Skip to content

This repository builds a product detection model to recognize products from grocery shelf images.

License

Notifications You must be signed in to change notification settings

sayakpaul/Grocery-Product-Detection

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Grocery-Product-Detection

This repository builds a product detection model to recognize products from grocery shelf images. The dataset comes from here. Everything from data preparation to model training is done using Colab Notebooks so that no setup is required locally. All the relevant commentaries have been included inside the Colab Notebooks.

Repository organization

├── Colabs
│   ├── GroceryDataset_EDA_Prep.ipynb: EDA and data preparation notebook. 
│   ├── GroceryDataset_Evaluation.ipynb: Runs evaluation on the test images with the trained model.
│   ├── GroceryDataset_Inference.ipynb: Performs inference with the trained model.
│   └── GroceryDataset_Model_Training.ipynb: Trains an SSD MobileDet model using TFOD API.
├── Deliverables
│   ├── image2products.json: Contains test image names as keys and the number of products contained in each image as values.
│   └── metrics.json: mAP, precision and recall computed on test set.
├── Misc Files
│   ├── confusion_matrix.csv: Confusion matrix computed on the test set using the trained model.
│   ├── generate_tfrecord.py: Generates TFRecords from the provided dataset. 
│   └── ssdlite_mobiledet_dsp_320x320_products_sync_4x4.config: Configuration file needed by the TFOD API. 
└── README.md

Results

Following snaps taken from TensorBoard after loading the evaluation logs (logs are available here) -

As we can see with 10k training steps the metrics keep on shining. I believe with more sophisticated hyperparameter tuning and a longer training schedule performance can further be improved.

Notes

  • The provided dataset is converted to TFRecords for easy compatibility with the TFOD API. Further notes are available inside the Colabs/GroceryDataset_EDA_Prep.ipynb notebook.
  • Augmentation used
    • Horizontal flips
    • Random crops
  • Detection network used: SSD MobileDet.
  • Training hyperparameters are available inside Misc\ Files/ssdlite_mobiledet_dsp_320x320_products_sync_4x4.config file.
  • Precision and recall reported in Deliverables/metrics.json are mean values computed over the [email protected] and [email protected] columns of Misc\ Files/confusion_matrix.csv.
  • A threshold of 0.6 was used in order to obtain the number of products per test image.

Trained model files

Find them here. If you are looking for the checkpoints, the latest ones are prefixed with model.ckpt-10000. There's also a frozen inference graph.

Dataset citation

@article{varol16a,
      TITLE = {{Toward Retail Product Recognition on Grocery Shelves}},
      AUTHOR = {Varol, G{"u}l and Kuzu, Ridvan S.},
      JOURNAL = {ICIVC},
      YEAR = {2014}
}

About

This repository builds a product detection model to recognize products from grocery shelf images.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published