CORNE

# CORNE CORNE is the large-scale training dataset used by OSOR for effect-aware object removal. It is built from instruction-based image editing triplets using SAVP, the Semantic-Anchored Verification Pipeline released in the OSOR repository. Dataset page: https://modelscope.cn/datasets/ZhouqmR/CORNE ## Dataset Summary Real-world object removal requires supervision that covers not only the target object, but also object-associated effects such as cast shadows, reflections, and residual traces. CORNE provides paired object-present images, paired clean backgrounds, effect-aware masks, and object-core masks for training one-step object removal models. CORNE is produced by applying SAVP to single-edit add/remove samples. SAVP verifies that the image difference is localized and semantically aligned with the edit instruction, then fuses validated perceptual differences with SAM2 object-core masks to produce effect-aware masks. ## Statistics | Item | Count | | --- | ---: | | Source samples considered | 680,088 | | Add/remove source samples | 427,929 | | CORNE retained training pairs | 287,012 | | Phase II effect-heavy subset | 67,726 | The Phase II subset is intended for incomplete-mask conditioning, where an object-core mask is used as the input condition while the effect-aware mask remains the supervision target. ## Data Structure Files are organized by matching image stems. A typical release contains: ```text CORNE/ ├── shot/ # I_shot: object-present input images ├── bg/ # I_gt: paired clean background images ├── mask-check/ # effect-aware masks covering object and associated effects ├── mask_sam/ # object-core masks from SAM2 └── *.txt # optional list files for training scripts ``` For each sample id `<stem>`: ```text shot/<stem>.png bg/<stem>.png mask-check/<stem>.png mask_sam/<stem>.png ``` ## How To Use Use `shot` as the object-present input image, `bg` as the reconstruction target, and `mask-check` as the effect-aware training mask. For Phase II alpha-aware training, use `mask_sam` or perturbed object-core masks as incomplete conditioning masks while continuing to supervise with `mask-check`. In the OSOR training configs, `mask-check` corresponds to `mask_list` / the effect-aware target mask. The OSOR training code expects list files: ```text shot_list.txt bg_list.txt mask_list.txt mask_sam_list.txt ``` Each list contains one path per line, either absolute or relative to the configured `data_root`. ## Construction Pipeline CORNE is generated with SAVP: 1. Select single add/remove edit triplets. 2. Convert each triplet into an ordered removal pair `(I_shot, I_gt)`. 3. Compute a multi-feature perceptual difference mask. 4. Verify difference regions with GroundingDINO semantic boxes. 5. Segment object-core regions with SAM2. 6. Fuse validated difference regions and object-core masks into effect-aware masks. 7. Select effect-heavy cases for Phase II incomplete-mask conditioning. See the `SAVP/` folder in the OSOR repository for code and constants. ## Intended Use CORNE is intended for research on object removal, image inpainting, effect-aware editing, and robust mask-conditioned restoration. It is especially suited for training models that must remove target objects together with shadows, reflections, and other non-local effects.