CORNE-Val

CORNE-Val is the held-out validation benchmark associated with CORNE and OSOR. It contains paired object-removal samples with both object-core and effect-aware masks.

Dataset page: https://huggingface.co/datasets/QinmingZhou/CORNE-Val

Dataset Summary

CORNE-Val contains 219 held-out samples generated with the same SAVP pipeline used for CORNE. It is intended for evaluating object removal models on paired backgrounds and dual mask settings.

Dataset Structure

CORNE-Val/
├── shot/       # object-present input images
├── bg/         # paired clean background images
├── mask_sam/   # object-core masks
└── mask_eff/   # effect-aware masks

For each sample id <stem>:

shot/<stem>.png
bg/<stem>.png
mask_sam/<stem>.png
mask_eff/<stem>.png

Usage

Use shot as the input image and bg as the paired target. Evaluate under:

• mask_sam: conservative object-core masks.
• mask_eff: effect-aware masks covering the object and associated effects.

This split is useful for validating both reconstruction quality and robustness to conservative masks.

Dataset Creation

CORNE-Val is sampled from held-out NHR-Edit shards reserved from CORNE construction. SAVP filters instruction-based triplets for localized and semantically aligned changes, then synthesizes effect-aware masks and object-core masks.

Intended Use

CORNE-Val is intended for validation and benchmarking of object removal, inpainting, and mask-conditioned restoration methods.