This repository builds a product detection model to recognize products from grocery shelf images.
APACHE-2.0 License
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.
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
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.
Colabs/GroceryDataset_EDA_Prep.ipynb
notebook.Misc\ Files/ssdlite_mobiledet_dsp_320x320_products_sync_4x4.config
file.Deliverables/metrics.json
are mean values computed over the [email protected]
and [email protected]
columns of Misc\ Files/confusion_matrix.csv
.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.
@article{varol16a,
TITLE = {{Toward Retail Product Recognition on Grocery Shelves}},
AUTHOR = {Varol, G{"u}l and Kuzu, Ridvan S.},
JOURNAL = {ICIVC},
YEAR = {2014}
}