Susuji/SnapPhysics-Benchmark

SnapPhysics Mass Estimation Benchmark

This dataset provides a mass estimation benchmark for the 50 objects evaluated in the SnapPhysics paper, SnapPhysics: A Physics-Aware Scene Graph from a Single View for Interactive Mixed Reality Scenes. Three separate tables provide mass ground truth, baseline estimates, and SnapPhysics estimates for the same 50 benchmark objects, with one row per object in each table. Images, masks, and 3D reconstructions are supporting assets.

Mass data

The GT, baseline, and SnapPhysics mass result tables are grouped in Mass_Estimation_Result/.

File	Contents
GT_object_mass.csv	GT masses for the 50 benchmark objects, taken from the supplied category-specific GT JSON snapshots
baseline_mass_estimation.csv	VLM_only, sam3d, and sam3d+sg mean mass estimates for the same 50 objects
Snapphysics_mass_estimation.csv	SnapPhysics mean mass estimates, standard deviations, and image-specific center-of-gravity estimates for the same 50 objects
snapphysics_CoG_estimation.csv	Center-of-gravity estimates for the 96 linked object observations, with crop and PLY references
observations.csv	Links from evaluation objects to observations, scenes, masks, crops, and 3D files
scenes.json	Scene images, object file lists, coordinate frames, and source mappings
raw_estimates/	Individual baseline and SnapPhysics repeat estimates used to compute the summaries

All masses are in kilograms. Original image identifiers are recorded in scenes.json. GT source filenames and JSON paths are recorded in the tables.

For each object observation in an input image, each method was evaluated with 10 independent runs × 10 repetitions per run, for 100 inference attempts. The estimation tables report mean mass estimates. std_mass_kg is the population standard deviation across valid repeats. Individual prediction records are provided in raw_estimates/.

Estimated Center of Gravity

Estimated center-of-gravity coordinates are provided in snapphysics_CoG_estimation.csv, with separate 2D and 3D coordinate columns and one row per object observation.

These are estimates, not measured ground truth. Image coordinates are normalized within the referenced crop; 3D coordinates use the referenced PLY's original frame in arbitrary units, not meters.

For each source observation, coordinates are the arithmetic mean of ten run summaries, each based on ten valid predictions. Stored 2D and 3D coordinates are rounded to four and six decimal places, respectively, before averaging. Coordinates from different observations are kept separate.

Scenes

Images are titled Dining room 1–5, Office 1–5, Living 1–5, and Desk Item 1–3. The object count is the number of reconstructed objects included with each scene entry.

Image title	Objects	Image and reconstruction files
Dining room 1	11	Open scene
Dining room 2	11	Open scene
Dining room 3	9	Open scene
Dining room 4	9	Open scene
Dining room 5	6	Open scene
Office 1	5	Open scene
Office 2	7	Open scene
Office 3	7	Open scene
Office 4	2	Open scene
Office 5	4	Open scene
Living 1	3	Open scene
Living 2	4	Open scene
Living 3	4	Open scene
Living 4	4	Open scene
Living 5	3	Open scene
Desk Item 1	3	Open scene
Desk Item 2	4	Open scene
Desk Item 3	3	Open scene

scenes.json lists the image and object file paths.

Reconstruction files

Each scene folder contains its image and the masks/, crops/, and 3D_objects/ subfolders.

PLY files contain Gaussian vertices and retain their original coordinates. For entries with asset_relationship=same_image, load the scene's object files together.