IxorThink developed a model allowing Spott to retrieve duplicate and similar products from their product database.
The model takes as input nothing more than an image of the product and outputs images and id's of duplicate or similar products.
IxorThink uses Dhashes to retrieve duplicate images. We use VGG16 feature extraction and an annoy index to retrieve similar images.
We build a second model, a visual fashion recognition model. That models allows Spott to search a product in their product database given a street image of that product. The difficulty lays in filtering out the background to recognize the product. IxorThink used siamese networks and triplet loss function to develop and train a model to do the job.
We also developed a video-tracking model. An object, within a bounding box, is tracked during several video frames. The image of the bounding the box is used as input to the visual fashion recognition model.
All these models together, lead to a software application that enables tracking and recognizing a product, based upon an image within a video-frame.