Travor278/droid-crossview-seg

DROID Cross-View Segmentation (private preview)

这是一个从 DROID / RoboInter 操作数据中筛选并人工复核的真实机器人多视角分割数据集。给定一个 视角中的目标物体图像与 mask，任务是在同一时刻的另一个相机视角中分割同一物体。数据集保留 ego（wrist）与两个第三人称视角，并记录每个最终 mask 的来源及人工修订过程。

This dataset evaluates cross-view object correspondence in real robot manipulation: given an object mask in a source camera, segment the same physical object in a synchronised target camera. The source material is the DROID subset distributed through RoboInter; target identities and masks were independently proposed, reviewed, and curated in Label Studio.

---

1. At a glance

usable view pairs	1,396
valid ego + exo1 + exo2 pairs	1,081
target_not_visible exo1 + exo2 pairs	315
directed relation records	7,116
exported view images / masks	3,873 / 3,873
distinct source moments (source_key)	1,249
distinct trajectory IDs	1,024
distinct free-form target strings	799
export size	3.03 GiB
mask resolutions	RLE [H,W] = [376,672] (1 masks) · RLE [H,W] = [720,1280] (3,872 masks)

file	records	direction
ego2exo.json	2,162	wrist/ego → exo1 or exo2
exo2ego.json	2,162	exo1 or exo2 → wrist/ego
exo2exo.json	2,792	exo1 ↔ exo2

The distinction between counts matters. A Label Studio task is one reviewed three-view source moment and target identity; an accepted pair is one such task after status filtering; a directed relation record is one source-camera to target-camera evaluation direction. A valid pair yields six directed relations, whereas target_not_visible yields only the two exo↔exo directions.

---

2. Layout

ego2exo.json
exo2ego.json
exo2exo.json
data/<task_key>/ego/0.png
data/<task_key>/exo1/0.png
data/<task_key>/exo2/0.png
records.jsonl
issues.jsonl
audit.json
checksums.jsonl
labelstudio-snapshot.json
previews/

• task_key is an immutable curation identity, not a semantic class label.
• ego is the DROID wrist camera; exo1 and exo2 are synchronised external views.
• Every relation contains one selected moment rather than a continuous video clip.
• records.jsonl is the canonical per-pair mask/provenance ledger.
• labelstudio-snapshot.json is a frozen audit snapshot, not the recommended training interface.

---

3. Record format

The format follows jaychempan/Ego-Exo4D-Relation-Test.

{
  "<pair_id>": {
    "video_id": "<pair_id>",
    "video_path": ["data/<task_key>/<target_view>/0.png"],
    "prompt": {
      "first_frame_image": "data/<task_key>/<source_view>/0.png",
      "first_frame_anns": {
        "0": {
          "segmentation": {"counts": "<COCO RLE>", "size": [720, 1280]},
          "area": 45853.0,
          "category_id": 1.0,
          "text": "target description"
        }
      }
    },
    "objects": {
      "0": {
        "video_id": "<pair_id>",
        "obj_id": 0,
        "final_caption": "target description",
        "video_path": ["data/<task_key>/<target_view>/0.png"],
        "segmentation": {"counts": "<COCO RLE>", "size": [720, 1280]},
        "area": 61042.0
      }
    }
  }
}

field	meaning
prompt.first_frame_image	source-view image whose object mask is supplied
prompt.first_frame_anns.0.segmentation	source/query mask in compressed COCO RLE
objects.0.video_path[0]	synchronised target-view image
objects.0.segmentation	target-view ground-truth mask
text / caption fields	reviewer-approved free-form target description

Loading

import json
from pathlib import Path
from PIL import Image
from pycocotools import mask as coco_mask

root = Path(".")
records = json.loads((root / "exo2exo.json").read_text())
record = next(iter(records.values()))

source = Image.open(root / record["prompt"]["first_frame_image"])
rle = record["prompt"]["first_frame_anns"]["0"]["segmentation"]
mask = coco_mask.decode({
    "counts": rle["counts"].encode("ascii"),
    "size": rle["size"],
})
assert mask.shape == tuple(rle["size"])

Always use the RLE size field. Most masks are native 720×1280, but the export keeps source-native dimensions and does not assume that every frame is identical.

---

4. Selection and mask provenance

Reviewer status controls the exported views:

• valid: export wrist/ego, exo1, and exo2;
• target_not_visible: omit wrist/ego and export exo1 + exo2 only;
• bad_sync_or_source or reject: exclude the pair.

The final mask is resolved independently for each view:

1. checked Independent Codex + SAM3 → use the independent SAM3 candidate;
2. checked RoboInter reference → use the upstream non-independent fallback;
3. neither checked → use the submitted human annotation;
4. new human regions drawn over a selected/copied base are combined by pixelwise union;
5. an edited region retaining the prediction's region ID replaces that base, so intentional erasures are preserved rather than re-added.

provenance	view masks	share
independent_sam3	3,234	83.5%
submitted_annotation_union	413	10.7%
official_fallback	213	5.5%
independent_sam3 + human_addition	13	0.3%

official_fallback counts output view masks; audit.json separately records 204 unique source/view fallback identities loaded by the exporter.

---

5. Mask and label statistics

There are 3,873 exported masks. Area range: 349 → 755,027 px, median 45,650 px, mean 92,252 px.

mask area	masks	share
< 1k px	33	0.9%
1k – 5k px	203	5.2%
5k – 20k px	682	17.6%
≥ 20k px	2,955	76.3%

Target strings are descriptive and intentionally not mapped to a closed ontology. The most frequent strings are: towel (24), white towel (22), cloth (21), white cloth (21), green towel (15), clothes (13), white pillow (13), pillow (12), green cloth (12), orange towel (11), paper towel roll (10), black cloth (10), yellow towel (10), lid (9), curtain (8).

---

6. Curation and quality control

1. Source sampling. Synchronised wrist/exo1/exo2 moments were sampled from the DROID portion of RoboInter with an emphasis on visual and trajectory diversity.
2. Target identity. A vision-language pass described one cross-view-consistent manipulated object in plain language.
3. Independent segmentation. SAM3 generated view-specific candidates from those descriptions and point/box prompts; these candidates were separate from the RoboInter reference masks shown for comparison.
4. Human review. Every task was reviewed in Label Studio. Reviewers could accept a candidate, accept the fallback, edit/extend a mask, rename the target, restrict a task to exo-only, or reject it.
5. Deterministic export. Ambiguous source selections, missing required masks, or active drafts stop export rather than being guessed.
6. Integrity checks. Every shipped file is listed in checksums.jsonl; mask RLEs, decoded areas, image dimensions, relation counts, and partial-file absence were independently verified.

The frozen review contained 1,500 submitted tasks: 1,396 accepted, 104 rejected/bad-source, 0 unresolved issues, and 0 active drafts. The export includes 8 four-panel union examples under previews/ and 3,891 checksum-covered files.

---

7. Known limitations

• No exact scene count is claimed. The frozen metadata contains 1,024 distinct trajectory IDs and 1,249 distinct source moments, but trajectory IDs are not scene IDs. They cannot be used to infer how many of DROID's collection scenes are represented.
• Single moments, not clips. Every camera contributes one synchronised frame per pair; video_path is schema-compatible but not a temporal sequence here.
• Free-form labels. Near-synonyms such as cloth, towel, and color-qualified variants remain separate strings.
• Model-assisted masks. SAM3 candidates were human-reviewed, but the dataset is not simulator-perfect ground truth. Provenance is supplied so experiments can filter sources.
• Native resolution varies. Decode each mask with its own [H, W]; do not assume square inputs or resize masks independently of their image.
• Curated subset. This is not the complete DROID or RoboInter release and should not be used to reproduce their headline scene/task statistics.

8. Attribution, terms, and status

Source projects:

• DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset
• RoboInter
• Ego-Exo4D Relation Test format

This derived private preview does not relicense upstream imagery or annotations. Use is subject to the applicable DROID and RoboInter terms. The curation/export code and this dataset card do not override those terms.

Status: private preview. The relation schema is stable; counts may change only through a newly versioned, fully audited export.