paingoat/sme-add-remove

Spatial Markup Editing — Add/Remove

Private. Non-commercial research use only. Derived from OBER, which is released under the NTU S-Lab License 1.0. Do not make this repository public or redistribute its contents.

Image-editing pairs where the edit target is indicated by a spatial markup drawn on the input image.

Operation	Samples	Prompt	Subsets
remove	200	remove it	captured 150, synthetic_shadow_single_object 40, synthetic_reflection 10
add	199	add {object}	captured 149, synthetic_shadow_single_object 40, synthetic_reflection 10

add is built by swapping a remove pair, so its supervision is approximate: the ground truth can contain shadows or reflections the prompt does not mention. One add sample (ober-cap-07746) is missing because the VLM caption ("stone") was rejected as attribute-only.

Layout

manifests/   all|remove|add .jsonl/.csv, qa_report.json, qa_grid_*.jpg
processed/   remove/<sample_id>/ and add/<sample_id>/:
             input.png gt.png markup.png input_markup.png preview.png metadata.json

Each manifest row has sample_id, paired_sample_id, operation, prompt, object, file paths relative to the repository root (input, gt, markup, input_markup, preview, metadata), width, height, bbox_xyxy, subset, split, global_row and qa_status (pass / warn; every warn is markup_clipped).

Loading

from huggingface_hub import snapshot_download
from datasets import load_dataset

root = snapshot_download("paingoat/sme-add-remove", repo_type="dataset")
ds = load_dataset("json", data_files=f"{root}/manifests/all.jsonl")["train"]
# image paths in each row are relative to `root`

Provenance

Objects captioned with Qwen/Qwen3-VL-2B-Instruct (prompt_v1). Source: OBER, pinned revision 25e064b39bf675fefb9ac4d2e2cd962037b7206c.