edges2shoes

edges2shoes is a dataset and task used in image-to-image translation research, focusing on converting edge-based representations into photorealistic shoe images. It was introduced as part of the pix2pix framework to illustrate conditional adversarial networks’ ability to translate one domain into another. The dataset comprises paired samples: an edge map or sketch of a shoe and the corresponding real photograph of that shoe. These pairs enable supervised training of models that learn to map simple outlines to detailed color images.

In typical setups, a conditional generative adversarial network is trained with a U-Net style generator and a PatchGAN discriminator. The training objective combines an adversarial loss, which encourages realistic outputs, with a reconstruction term (often L1) that preserves overall structure and alignment with the input edge map. Edge maps in the dataset are generated either by edge detection techniques or manual drawing, and each edge is paired with the exact same shoe image to provide a ground truth target.

Uses and evaluations of edges2shoes center on demonstrating image-to-image translation capabilities and benchmarking model performance on paired data. Researchers assess quality through qualitative visual inspection and metrics such as reconstruction error, along with standard GAN-based assessments. The dataset has contributed to understanding how well models can infer texture, color, and shading from simple outlines.

Limitations include the dataset’s domain specificity to shoes, potential biases in photography and styling, and the inherent ambiguity in translating line drawings to full-color images. Edges2shoes remains a common reference point for exploring conditional generation and bridging abstract sketches to realistic object imagery.