TYTSTQ/ordinary-bench-grid3d

ORDINARY-BENCH Grid3D Dataset

A benchmark dataset for evaluating Vision-Language Models (VLMs) on 3D object localization in a 4x4x4 grid using 6 orthographic views.

Source code & evaluation pipeline: GitHub - tasd12-ty/ordinary-bench-core

Related dataset (free-placement scenes): TYTSTQ/ordinary-bench

Overview

Scenes	140 synthetic 3D scenes (Blender, 4x4x4 grid)
Complexity	7 levels: 4 to 10 objects per scene (20 each)
Views	6 orthographic projections per scene (480x480 PNG)
Task	Determine each object's grid cell from the 6 views

Task Description

Objects are placed in a 4x4x4 discrete grid (64 possible positions). The VLM receives 6 labeled orthographic views and must output each object's grid cell position.

Coordinate System

Row:    A, B, C, D    (A = front, D = back)
Column: 1, 2, 3, 4    (1 = left,  4 = right)
Layer:  1, 2, 3, 4    (1 = bottom, 4 = top)

Position format: RowCol-Layer
Example: "B3-4" = Row B, Column 3, Layer 4

View Projections

View	What it shows	Axes
Top (looking down)	Row + Column	Rows A-D top-to-bottom, Cols 1-4 left-to-right
Front (from front)	Column + Layer	Cols 1-4 left-to-right, Layers 1-4 bottom-to-top
Right (from right)	Row + Layer	Rows A-D left-to-right, Layers 1-4 bottom-to-top
Back (from back)	Column + Layer	Cols 4-1 left-to-right (reversed), Layers 1-4 bottom-to-top
Left (from left)	Row + Layer	Rows D-A left-to-right (reversed), Layers 1-4 bottom-to-top
Bottom (looking up)	Row + Column	Rows A-D top-to-bottom, Cols 4-1 left-to-right (reversed)

Any two orthogonal views (e.g., Top + Front) are sufficient to determine all 3 coordinates, but 6 views provide redundancy for verification.

Quick Start

from datasets import load_dataset

ds = load_dataset("TYTSTQ/ordinary-bench-grid3d", split="test")

sample = ds[0]
sample["image_top"]       # PIL Image (480x480) - top-down view
sample["image_front"]     # PIL Image (480x480) - front view
sample["system_prompt"]   # System prompt with coordinate system
sample["user_prompt"]     # User prompt with view labels + object list
sample["ground_truth"]    # JSON: [{"object": "cyan rubber cylinder", "cell": "B3-4"}, ...]

Data Splits

Split	Scenes per complexity	Total scenes
train	15	105
test	5	35

Column Schema

Column	Type	Description
scene_id	string	Scene identifier, e.g., g07_000010
n_objects	int	Number of objects (4-10)
split	string	Complexity split: g04 through g10
image_top	Image	Top orthographic view (looking down)
image_bottom	Image	Bottom orthographic view (looking up)
image_front	Image	Front orthographic view
image_back	Image	Back orthographic view
image_left	Image	Left orthographic view
image_right	Image	Right orthographic view
objects	string	JSON: object descriptions
system_prompt	string	System prompt for VLM
user_prompt	string	User prompt with view annotations
ground_truth	string	JSON: [{"object": "...", "cell": "B3-4"}, ...]
scene_metadata	string	Full scene JSON (3D coords, grid info)

Expected Response Format

[
  {"object": "cyan rubber cylinder", "cell": "B3-4"},
  {"object": "brown metal sphere", "cell": "A1-2"}
]

Scoring Criteria

Metric	Description
Exact accuracy	Predicted cell == GT cell
Structural accuracy	Correct under D4 symmetry (rotation/reflection in row-col plane)
Per-dimension accuracy	Row, Column, Layer accuracy independently

The D4 symmetry metric accounts for consistent rotations/reflections of the entire grid, which can happen when VLMs misinterpret view orientations.

Source Code

Full generation pipeline, VLM evaluation, and scoring tools:

github.com/tasd12-ty/ordinary-bench-core/tree/main/data-gen-grid3d

License

MIT