Vivid3Dai/kitchen2-edge-cases-synthetic-data

Home Robot View: Indoor Objects and Edge Cases

Synthetic indoor visual data from Vivid 3D. Robot-view sequences in a second residential interior, with furniture, people, lemons, a dog, and a broken vase. Scene metadata includes robot states such as picking up an object and observing an event.

Dataset contents

Item	Count or format
Images	1,200
Image resolution	1920 × 1080 pixels
Render iterations	5, with 240 frames each
Object annotations	8,970
Source binary masks	8,970, one file per object instance
Source category IDs	20
Depth and LiDAR	1,200 EXR and 1,200 NPZ files in the original archive
Scene metadata	1,200 JSON files

COCO category IDs run from 1 to 20. Each category also retains its original Vivid resource ID as source_class_id. Names are taken from the source stats. Distinct furniture and person resources remain distinct categories. See classes.csv for the mapping. These are dataset-specific categories, not the standard COCO label set.

Download and use

The Parquet files provide one row per image, with the image, instance masks, and annotation metadata.

• COCO package: images and annotations/instances_train.json, with boxes and instance segmentation.
• COCO annotations: the annotation JSON separately.
• Original export: source RGB, YOLO, masks, depth, LiDAR, and scene metadata without changes.

from datasets import load_dataset

dataset = load_dataset("Vivid3Dai/kitchen2-edge-cases-synthetic-data")
sample = dataset["train"][0]
image = sample["image"]
masks = sample["masks"]  # One binary image for each annotated instance

All 1,200 images are included. The train split name is used for loading convenience. No independent validation or test split was supplied. These are sequential frames. Keep related sequences together when creating evaluation splits.

Viewer fields

Fields	Meaning
image, masks	Original encoded RGB image and a list of original binary instance masks
image_id, coco_image_id, source_image_path	Source frame identifier, COCO image ID, and original image path
width, height, iteration	Image dimensions and source render iteration
source_class_ids, bboxes_yolo, bboxes_source_xywh	Source IDs and boxes in the original YOLO row order. YOLO coordinates are normalized. bboxes_source_xywh converts them to pixels
mask_class_ids, category_ids, annotation_ids	Original mask resource IDs, COCO category IDs, and COCO annotation IDs, in mask order
bboxes_xywh, areas, source_mask_paths	Pixel boxes and foreground pixel areas derived from masks, plus original mask paths, in mask order
source_detection_text, source_stats_json	Original detection text and scene metadata

masks and all fields described as being in mask order align with one another. The source YOLO lists form a separate group. Do not pair a YOLO row with a mask only by list index.

Annotation conventions

• Each source object mask is encoded as COCO RLE. COCO boxes use pixel [x, y, width, height], with a top-left origin. Boxes and pixel areas are derived from the masks. The original normalized YOLO boxes remain available separately.

• Several objects can share one source category ID, such as multiple lemons. They remain separate instances. Mask filename suffixes such as (1) are not additional classes or verified persistent tracking IDs.

• The source contains 5 zero-width or zero-height YOLO boxes. Their nonempty masks contain thin, one-pixel-wide or one-pixel-high fragments. Mask-derived COCO boxes retain these visible fragments.

• Labels identify objects, not incident or defect classes. For example, broken vase fragments are still one Vase Simulate instance. A sink mask does not separately label water spray.

• Twenty annotations use raw ID 0. The corresponding entry in the source stats is Deco table, vividID 2795. The export retains source_class_id=0 and assigns it COCO category_id=1, named source_class_0. The metadata candidate is recorded separately. This is not a global class-index offset.

Source conventions and limitations

• Source RGB names use .png, although the encoded images are JPEG. Images in the COCO ZIP use .jpg names. Source mask files are PNG. Original files remain in the archive without changes.
• Depth EXR and LiDAR NPZ are supplied as additional source data. This card does not establish their physical units or sensor calibration.
• Scene JSON includes object identities, transforms, and scenario information. Robot state names are not joint-control action vectors or a ready-to-use robot policy dataset.
• The data is synthetic. No real-world model performance or defect-recognition result is claimed.

Source details and archive checksums are in source_manifest.json. Package file checksums are in SHA256SUMS.txt.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). Credit Vivid 3D when using or sharing this dataset. See LICENSE.

About Vivid 3D

Vivid 3D creates synthetic visual data from 3D content for computer vision workflows.