How are you representing clothing segments for cosine similarity—are you embedding the full segmentation masks, extracted features from a vision model (e.g., CLIP), or using texture/color histograms?