Samrt Fashion Project

Target

• Build a system to detect and segment the certain garment in the images (Link)

• Classification the product attribute-based fashion product (Link)

• Use Multimodal method - image and text to search fashion style (Link)

1. Detection&Segmentation, Multi-Labels Classification and Attributes Embedding

(Link)

• Detection and Segmentation Reseults:

• Multilabel Classification Results:

Multimodal Fashion style retrival

Link

• Methods:
  • Image Features: ResNet18 AvgPool2D feature layer: 512,
  • Textula Features: LSTM text encoding: hidden dim: 512,
  • Joint embedding: concatenate two vectors, concat(, )
  • Train Network: 2 layer MLPs with RELU, with batch-norm and dropout(0.1)
  • Loss Function: mini-batch retreival loss (paired images)
  • Metric: Recall@Top-K(1,5,10,50,100)
• Test Query Format:
  • Source Image:
  • Query Format: Change One Attribues
  • Return: Target Images
  • Example:
    • Source Image: women/pants/wide-leg_and_palazzo_pants/91340759/91340759_2.jpeg
    • Query: change black to green
• Query Demo