Testing-Dataset

This repository contains two complementary datasets used to evaluate the NON-GON collision-detection algorithms.

Random dataset — geometric parameters (position, size, rotation) are sampled randomly for each contact pair. The Gilbert–Johnson–Keerthi (GJK) algorithm and Expanding Polytope Algorithm (EPA) provide ground-truth collision status and signed distance. For full generation details see:

• PQ_DATASET_GENERATION.md for PQ dataset details
• SD_DATASET_GENERATION.md for SD dataset details

Edge-case dataset — hand-crafted cases designed to stress-test the algorithms against specific geometric configurations known to be difficult: near-boundary contacts, degenerate shapes, rotation extremes, and scale mismatches. See edge/EDGE_DATASET.md for full details.

Both datasets use GJK/EPA (via coal) as ground truth and serve as a high-fidelity reference benchmark within the project.

Dataset Summary

Dataset	Pair	Samples	Colliding	Non-colliding	Labels
PQ	Ellipse–Ellipse	2000	1000	1000	collision
PQ	Cylinder–Cylinder	2000	1000	1000	collision
PQ	Ellipsoid–Ellipsoid	2000	1000	1000	collision
PQ	Ellipsoid–Elliptic Paraboloid	2000	1000	1000	collision
PQ	Hyperboloid–Plane	2000	1000	1000	collision
SD	ConvexCircle–ConvexCircle	2000	1000	1000	collision, signed distance
SD	ConvexLine–Line	2000	1000	1000	collision, signed distance
SD	Ellipse–Ellipse	2000	1000	1000	collision, signed distance
SD	Ellipsoid–Ellipsoid	2000	1000	1000	collision, signed distance
SD	HemiEllipsoid–Plane	2000	1000	1000	collision, signed distance
SD	Point–Ellipse	2000	1000	1000	collision, signed distance
SD	Point–Ellipsoid	2000	1000	1000	collision, signed distance
SD	Superellipse–Line Segment	2000	1000	1000	collision, signed distance
SD	SuperEllipsoid–Plane	2000	1000	1000	collision, signed distance
Total	14 pairs	28000	14000	14000

Results

PQ Dataset — Collision Classification Accuracy

Pair	Correct	Total	Accuracy
Ellipse–Ellipse	1993	2000	99.65%
Cylinder–Cylinder	1937	2000	96.85%
Ellipsoid–Ellipsoid	1995	2000	99.75%
Ellipsoid–Elliptic Paraboloid	1754	2000	87.70%
Hyperboloid–Plane	1111	2000	55.55%

SD Dataset — Signed Distance RMSE

Pair	Samples	RMSE
ConvexCircle–ConvexCircle	2000	0.342734
ConvexLine–Line	2000	0.700965
Ellipse–Ellipse	2000	0.139106
Ellipsoid–Ellipsoid	2000	0.966192
HemiEllipsoid–Plane	2000	4.887331
Point–Ellipse	2000	0.007352
Point–Ellipsoid	2000	0.723910
Superellipse–Line Segment	2000	0.010067
SuperEllipsoid–Plane	2000	2.281848

Edge-Case Dataset

Hand-crafted cases that target specific geometric configurations known to be difficult. Every case has a GJK/coal ground truth. See edge/EDGE_DATASET.md for full row-by-row breakdowns and failure analysis.

Structure

edge/
├── PQ/   — Proximity Query (binary collision label)
│   ├── EllipseEllipse.csv                (30 rows)
│   ├── EllipsoidEllipsoid.csv            (30 rows)
│   ├── CylinderCylinder.csv              (30 rows)
│   ├── EllipsoidEllipticParaboloid.csv   (30 rows)
│   └── HyperboloidPlane.csv              (36 rows)
└── SD/   — Shortest Distance (signed distance value)
    ├── EllipseEllipse.csv                (30 rows)
    ├── EllipsoidEllipsoid.csv            (30 rows)
    ├── ConvexCircleCircle.csv            (30 rows)
    ├── ConvexLineLine.csv                (30 rows)
    ├── PointEllipse.csv                  (30 rows)
    ├── PointEllipsoid.csv                (30 rows)
    ├── SuperellipseLineSegment.csv       (31 rows)
    ├── SuperEllipsoidPlane.csv           (30 rows)
    └── HemiEllipsoidPlane.csv            (30 rows)

PQ Edge Results

Pair	Pass	Total	Notes
EllipseEllipse	30	30	All categories pass
EllipsoidEllipsoid	30	30	All categories pass
CylinderCylinder	30	30	All categories pass
EllipsoidEllipticParaboloid	29	30	False positive at gap=0.001 from apex (row 15)
HyperboloidPlane	23	36	Fails for plane in upper half of height range, small rotations (10°–30°), and near-degenerate shapes

SD Edge Results

Failures on overlapping cases (collision=1, GT < 0) are expected: the alternating-projection algorithm returns surface-to-surface distance, not penetration depth. Only failures on separated cases (GT > 0) represent genuine bugs.

Pair	Pass	Total	Genuine bugs (separated cases)
EllipseEllipse	18	30	None — all failures are overlap cases
EllipsoidEllipsoid	17	30	None — all failures are overlap cases
ConvexCircleCircle	20	30	Rows 26–27: direction-dependent error (y-axis and 45° separations wrong)
ConvexLineLine	16	30	Multiple rows: y-axis projection only — ignores x-offset, returns 0 for horizontal exterior
PointEllipse	24	30	Rows 26, 29: 89° rotation and near-circle
PointEllipsoid	28	30	Row 28: diagonal approach, error exactly +2.0
SuperellipseLineSegment	25	31	None — all failures are overlap cases
SuperEllipsoidPlane	24	30	Rows 11, 18–22: error grows with rotation 30°→75°, vanishes at 90°
HemiEllipsoidPlane	29	30	None — single failure is overlap case

Genuine Algorithm Bugs

Pair	Trigger
HyperboloidPlane PQ	Plane in upper half of height range; small rotations (10°–30°); near-degenerate shapes (c→0, c→∞, asymmetric axes)
EllipsoidEllipticParaboloid PQ	False positive at gap=0.001 from apex
SuperEllipsoidPlane SD	Error grows with rotation 30°→75°, drops to ~0 at 90° (principal-axis projection)
PointEllipsoid SD	Diagonal approach — error exactly +2.0 (axis-projection bug)
PointEllipse SD	89° rotation and near-circle
ConvexLineLine SD	y-axis projection only — returns 0 for horizontal exterior; ignores x-offset
ConvexCircleCircle SD	Direction-dependent: correct on x-axis, wrong on y-axis and 45°