outGANfit
cDCGANs‐based Architecture to Generate Outfit Items Compatible with the Input Garment
outGANfit (code available here) is a conditional Deep Convolutional Generative Adversarial Networks (cDCGANs‐)based architecture to generate outfit items compatible with the input garment, using the Polyvore Outfits dataset.
The dataset
The dataset, Polyvore Outfits (downloadable at this link) consists of 261,057 RGB images of garments measuring 300x300 pixels, divided into the categories “tops”, “bottoms”, “shoes”, “accessories”, “all-body”, “jewellery”, “bags”, “hats” and “outerwear”. From these, only the first four were chosen. The goal is to input a t-shirt to the model and have it generate a matching pant, pair of shoes and accessory (only sunglasses in the available implementation).
outGANfit architecture
The architecture includes three different GANs, one for each garment to be predicted (pants, shoes, accessories). For each gan there is a generator and two discriminators, one of which is used to evaluate the compatibility between the conditioning image (t-shirt) and the generated garment. The other discriminator is useful, on the other hand, to classify what is generated as real or fake.
Results
The architecture has been tested - obtaining a FID value of 175.4166. Below, some outputs:
