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.
- 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 ofMisc\ Files/confusion_matrix.csv
. - A threshold of 0.6 was used in order to obtain the number of products per test image.
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}
}