Repository files navigation

Fashion200K dataset used in ICCV'17 paper "Automatic Spatially-aware Fashion Concept Discovery."[paper]

Author: Xintong Han

Contact:xintong@umd.edu

This dataset is crawled fromLyst.com in September 2016.

You can download the dataset viaGoogle Drive:

image_urls.txt: Image urls. Format: ImageName ImageURL.

detection: Detection results. Format: ImageName Category_DetectionScore_xMin_xMax_yMin_yMax. We trained aMultiBox detector for 9 classes (dress, skirt, top, bag, shorts, sunglasses, shoe, outwear, pants).

women.tar.gz: Cropped detected images. If you want the original images, they can be downloaded using their urls in image_urls.txt.

labels: train/test labels. Format: ImageName DetectionScore ProductDescription.

Note that there is information of more than 300k images in image_urls.txt and detection folder, but we remove around 100k images (see labels folder) because they have low detection scores.

You can refer to Tensorflow'sim2txt for how to train the model. By setting f_rnn_loss_factor and g_rnn_loss_factor to 0 in the model configuration in thisrepo can also train a visual-semantic embedding for the same purpose.

You can look at the originalCAM code to figure out how to extract activation maps in a joint embedding setting.

@inproceedings{han2017automatic,  title = {Automatic Spatially-aware Fashion Concept Discovery},  author = {Han, Xintong and Wu, Zuxuan and Huang, Phoenix X. and Zhang, Xiao and Zhu, Menglong and Li, Yuan and Zhao, Yang  and Davis, Larry S.},  booktitle = {ICCV},  year  = {2017},}