FrameCut3D

FrameCut3D is a benchmark for out-of-frame truncation in single-image 3D generators: the object in the input photo runs off the image frame, and the generator silently returns a truncated shape. It accompanies the paper "FrameCut3D: Diagnosing Out-of-Frame Completion in Single-Image 3D Generators" (under review).

Real split (66 objects, real/)

Everyday objects captured with a handheld phone orbit. Each object folder contains:

File	Description
object.png	Isolated full-object image (background removed), the complete input.
cuts/cut40.png ... cut70.png	Controlled truncation: the bottom 40 to 70% of the object's bounding box is masked.
frame.jpg	The full capture frame the object image was taken from.
crop.png, crop.json	Genuine frame truncation: the fixed center crop of the capture, with its severity and cut sides (6 objects are not cut).
measured_geometry.ply	Reference geometry from structure-from-motion and multi-view stereo on the orbit, support plane removed (coloured points).
pseudo_gt.ply	The generator's reconstruction from the complete image (TripoSplat Gaussians), used as pseudo ground truth.
video.mp4	The capture orbit (1080p, 30 fps).
meta.json	Label, description, category, geometry tier (1 solid to 4 reflective or transparent), symmetry flag, common/uncommon category flag.

The measured geometry misses part of the surface next to the support plane. The paper quantifies this bias. Coordinates are in the reconstruction frame (arbitrary scale and rotation); evaluation normalizes each shape to a unit box and searches the rotation.

Synthetic splits (synthetic/)

• objaverse_48.json: the 48 Objaverse objects of the main synthetic split (UIDs, one view at azimuth 180, elevation 20). Objaverse assets keep their original licences and are not redistributed here.
• gso_200.json: the 200 Google Scanned Objects of the held-out split (three views at azimuth 0, 120, 240).

Scripts (scripts/)

truncate.py (cut rule and hidden band) and bandmetrics.py (Chamfer, F-score, hidden-band recall after alignment).

Load

from datasets import load_dataset
ds = load_dataset("issai/FrameCut3D", split="train")
# image (complete), cut40 ... cut70, frame, crop, annotations, and paths to measured_geometry, pseudo_gt, video

Licence

CC BY-NC 4.0 for the real split and the annotations. Synthetic objects keep their source licences.

Citation

The paper is under review. A citation will be added on publication.