pubkey / fashion-segmentation
Detect which pixels of your product images actually belong to the product. We use up to 4 images as input because on many fashion images, you need multiple images as context to know which fashion item must be segmentated. In the following image you can see that the person wears a jacked, a t-shirt and a jeans. But only with all 4 images, it is clear to the network that the jacked must be segmentated.
The segmentation data enables you to extract the actual meaning of your product images by calculating the accurate color or extracting patches and shapes.
With the generated data, you can build next-level features, such as a detailed color search, attractive variation previews or find other products with the same color.
In this repo you can find the low resolution (128x128) model that can be used for testing it out. To get access to the better trained model with higher resolution input images, you can buy it from me. The higher resolution model gives a more detailed output and is able to better detect the shapes and surface of the fashion item which reduces false positives.
