- We have crawled sneaker images and metadata from StockX (https://stockx.com/), a leading online resale shop at global scale.
- Total number of sneaker products crawled: 23,492 (see Table 1 for basic statistics).
- Due to the copyright issue, we cannot provide the original sneaker images directly; instead, we share downloadable links as metadata.
- We share the data for 1) our embedding results, 2) SOTA embedding, and 3) various engineered features along with 4) the crawled metadata containing image downloadable links and 5) the entire history of resale transactions for the sample sneaker products from 2012 to 2020.
- See Section 2 for the details of each dataset provided via the link.
- See Section 3 for the details of the codes to reproduce the research outcomes.
- See section 4 for additional results not shown in the paper due to page limitations.

A. df_512_mask_shape_210927.csv (259.8MB):
- Color embedding: Constructed embedding for the latent shape-invariant representation (top module in Figure 1).
- Columns: 1) 0 ~ 511: 512 dimensional (D) embedding features; 2) 512: product ID; 3) 513 ~ 612: 100D metadata features.

B. df_512_mask_color_210927.csv (261.2MB):
- Shape embedding: Constructed embedding for the latent color-invariant representation (middle module in Figure 1).
- Columns: 1) 0 ~ 511: 512D embedding features; 2) 512: product ID; 3) 513 ~ 612: 100D metadata features.

C. df_512_mask_all_210927.csv (260.5MB):
- Combined embedding: Constructed embedding for the latent all-invariant representation (bottom module in Figure 1).
- Columns: 1) 0 ~ 511: 512D embedding features; 2) 512: product ID; 3) 513 ~ 612: 100D metadata features.

D. df_384_looc.csv (211.9MB):
- Constructed embedding for the Leave-one-out Contrastive Learning (LooC).
- We consider this one as SOTA and compare our model with this one (Xiao et al, ICLR 2021).
- Columns: 1) 0 ~ 383: 384D embedding features; 2) 384: product ID; 3) 385 ~ 484: 100D metadata features.

E. total_df_RGBHSV_ent_seg_rgbHistBin128_meta.pkl (108.6MB):
- Features from the feature engineering (see the manuscript for more detailed description).
- Columns: 1) 0: product ID; 2) 1 ~ 12: 12D color distribution parameters (mean, std) for RGB, HSV; 3) 13 ~ 19: 7D color entropy features for RGB, HSV, Grayscale; 4) 20 ~ 24: 5D image segmentation features; 5) 25 ~ 408: 384D color histogram of 128 bins for RGB; 6) 409 ~ 508: 100D metadata features.

F. val_sneakers_df.pkl (42.8MB):
- Metadata crawled from StockX.com.
- We share downloadable links ("imageUrl") for the sneaker image samples of the current study: see "crawling_source.ipynb" to download the images via the links.
- Columns: 1) 0: pid (product ID), images can be distinguished by this ID; 2) 1 ~ 100: a total of 100 metadata features, e.g., product name, retail price, colorway, release date, primary category, consumer type, and so on; 3) 101: imageUrl, downloadable links for the target sneaker images.

G. resale_transactions_ALL.csv (1.68GB):
- Containing ALL resale transactions for the target sneaker products, from 2012Q2 to 2020Q3.
- Columns: 1) quarter_sold; 2) month_sold; 3) original_resalePrice; 4) inflated_resalePrice: we compute adjusted (inflated) retail or resale price by following the process described in Figure 2; 5) original_retailPrice; 6) inflated_retailPrice; 7) original_profit; 8) inflated_profit, 9) date_of_release: sneaker release date; 10) image_fileName: pid.

A. crawling_source.ipynb (see in embSneakers/code/ in the repository):
- From the code, you can automatically download the sneaker images that have been sampled for the current work, via the links from the metadata: see the last column "imageUrl" in "val_sneakers_df.pkl."
- i.e., One image per sneaker product is downloaded and used in the current study.
- You only need to run "Crawling by imageUrl" part if you properly download "val_sneakers_df.pkl"; "RAW Crawling source (Using Selenium, Chromdriver)" part is used only when crawling from the scratch.

B. sneaker_embedding_masked_three_heads.ipynb (see in embSneakers/code/ in the repository):
- You need to download the sneaker image samples from A. first to construct embeddings from the given code.
- Otherwise, please refer to the code to understand how the embeddings are constructed in details.

C. classification_and_regression.ipynb (see in embSneakers/code/ in the repository):
- From the code, you can run classification and regression tasks by inputting the provided embedding datasets (three from our model, one from SOTA).
- Kinds of inference tasks: 1) Classifications for Primary Category (8 classes), Consumer Type (5 classes), and Maximum Resale Premium (2 classes); 2) Regression for Maximum Resale Premium.
- Kinds of inference model: 1) Multinomial Logistic Regression; 2) XGBoost; 3) Multi-layered Perceptron (MLP, a Neural-net method).