In this project, we have developed an object detection model for seven types of household waste: ['Drink can', 'Glass bottle', 'Other plastic bottle', 'Clear plastic bottle', 'Plastic bottle cap', 'Plastic film', 'Plastic straw']. We utilized the TACO dataset and the Kaggle dataset, along with the pretrained MobileNet V2 FPNLite 320x320 SSD model from the TensorFlow Model Zoo.
To preprocess the TACO and Kaggle datasets before using them, please refer to the step-by-step instructions outlined in Collect_Data.ipynb. However, if you wish to use your own dataset, please ensure that it is an image dataset with annotations available in either COCO JSON or Pascal VOC XML format. Also, remember to split your dataset into training and testing sets.
To utilize our object detection model for household waste in Google Colaboratory, please follow the steps provided in Fine_tuning_ssd_mobilenet_fpnlite_320.ipynb. To evaluate the model and obtain the mean Average Precision (mAP) metric, please refer to the instructions in Eval_ssd_mobilenet_fpnlite_320.ipynb.
To serve the model we use TensorFlow Serving with Docker. For more instructions head over to serving branch on this repo