ihaveamoose/DLCR

DLCR: A Generative Data Expansion Framework via Diffusion for Clothes-Changing Person Re-ID

Publicly hosted repository for the generated data presented in "DLCR: A Generative Data Expansion Framework via Diffusion for Clothes-Changing Person Re-ID", under review in WACV 2025 Algorithms Track.

We generate and release over 2.1M synthetic images across 4 CC-ReID datasets, namely PRCC, CCVID, VC-Clothes, and LaST. We use diffusion inpainting to change the subject's clothing in an image, increasing each dataset's clothing diversity by over 10x and improving downstream CC-ReID performance. Upon acceptance, please see the paper for more details.

The generated datasets maintain the same directory structure as the corresponding original datasets. For each original image, we created a .png file that includes all 10 variants of the original image. For example, given the original sample:

We store its 10 variants as a single image of the following form:

Additionally, we store a .txt file that contains the descriptions of the clothing for each variant, with each description on a separate line.

Gray shirt| Blue jeans| Black shoes
Black top| Black pants| Black shoes
White top| Blue jeans| White shoes
blue shirt| black shorts| black shoes.
Shirt | Jeans | Shoes
Yellow t-shirt| Black pants| Black shoes
Gray top|Blue jeans|White shoes
Green top| Blue shorts| White shoes
Gray t-shirt| Black pants| Black shoes
Black shirt| Black pants| Black shoes

To read the generated images when you have acces to the name of the original image, you can use the following code:

generated_images = os.path.join(self.generated_data_location, img_name+".png")
generated_images = Image.open(generated_images)
generated_images = np.array(generated_images)
generated_images = rearrange(generated_images, 'h (b w) c->b h w c', b=10)