-
Download images:
-
Download
img_highres.zip
from In-shop Clothes Retrieval Benchmark. You will need to follow the download instructions to unzip the file. Unzip file inDATASET/DeepFashion_highres/img_highres
-
Download the train/test data: train.lst, test.lst, and fashion-pairs-test.csv. Put in
DATASET/DeepFashion_highres/tools
. Note: because not all training images had their densepose detected we used a slightly modified training pairs file fashion-pairs-test.csv. -
Split the train/test dataset using:
python util/generate_fashion_datasets.py --dataroot DATASET/DeepFashion_highres
This will save the train/test images in
DeepFashion_highres/train
andDeepFashion_highres/test
.
-
-
Compute densepose:
-
Install detectron2 following their installation instructions.
-
Use apply net from densepose and save the train/test results to a pickle file. Make sure you download densepose_rcnn_R_101_FPN_DL_s1x.pkl.
-
Copy
util/pickle2png.py
into your detectron2 DensePose project directory. Using the DensePose environment, convert the pickle file to densepose png images and save the results inDATASET/densepose
directory, using:
python pickle2png.py --pickle_file train_output.pkl --save_path DATASET/densepose/train
-
-
Compute human foreground mask. Save results in
silhouette
directory. Or you can download our computed silhouettes for the training set and testing set. -
Compute UV space coordinates:
-
Compute UV space partial coordinates in the resolution 512x512.
- Download the UV space - 2D look up map and save it in
util
folder. - Compute partial coordinates:
python util/dp2coor.py --image_file DATASET/DeepFashion_highres/tools/train.lst --dp_path DATASET/densepose/train --save_path DATASET/partial_coordinates/train
- Download the UV space - 2D look up map and save it in
-
Complete the UV space coordinates offline, for faster training.
- Download the pretrained coordinate completion model from here.
- Complete the partial coordinates.
python util/complete_coor.py --dataroot DATASET/DeepFashion_highres --coordinates_path DATASET/partial_coordinates --image_file DATASET/DeepFashion_highres/tools/train.lst --phase train --save_path DATASET/complete_coordinates --pretrained_model /path/to/CCM_epoch50.pt
-
-
Download the following in
DATASET/resources
, to apply Face Identity loss:- download the pre-computed required transformation (T) to align and crop the face.
- Download sphereface net pretrained model.
Note: we provide the DeepFashion train/test split of StylePoseGAN [Sarkar et al. 2021]: train pairs, and test pairs.