Hotel ID classification to prevent human trafficking Uses the 2022 dataset with 49,703 images from 3116 classes. Following is the visualisation of a few samples from the dataset
First used pretrained VGG19 as base network in deep metric loss (triplet loss model). For this training, removed images from classes that contained less than 2 pictures.
After this training, removed final layer and added softmax to classify images. To decrease overfitting, data augmentation applied. Code in Tensorflow, Keras
Future changes to be made: Using softmax scores, output 5 possible predictions. Evaluate using Mean average precision. Try more techniques to prevent overfitting like dropout regularization Need to collect more real world data
Credits for model ideation: Stanford University Yuyu Lin, Peng Chen, Chi On Ho