Pose benchmark leaderboard

AVoptim

Calibration and trajectory refinement across four driving datasets.

Leaderboard

Calibration + trajectory refinement. Original and MOISST++ are compared across all ten pose-benchmark scores: rendering quality, structure from motion and geometric consistency.

Dataset averages over five sequences; overall averages over all twenty. Best scores in each metric are bold.
Dataset	Pose set	Nerfacto			Splatfacto			SfM		Geometry
Dataset	Pose set	PSNR dB ↑	SSIM ↑	LPIPS ↓	PSNR dB ↑	SSIM ↑	LPIPS ↓	Reprojection px ↓	Track length ↑	P2M precision % ↑	Surface distance m ↓
All datasetsMean over 20 sequences	OriginalDataset poses	21.80	0.720	0.455	23.11	0.772	0.333	1.108	6.11	54.98	0.296
All datasetsMean over 20 sequences	MOISST++Calibration + trajectory	22.69	0.760	0.387	25.03	0.820	0.266	0.886	6.95	68.74	0.204
KITTI-360Mean over 5 sequences	OriginalDataset poses	21.30	0.718	0.278	23.87	0.793	0.178	0.575	6.80	77.00	0.100
KITTI-360Mean over 5 sequences	MOISST++Calibration + trajectory	21.41	0.726	0.272	23.91	0.794	0.177	0.580	6.82	78.10	0.085
nuScenesMean over 5 sequences	OriginalDataset poses	20.72	0.708	0.531	21.88	0.747	0.404	1.524	6.13	45.60	0.352
nuScenesMean over 5 sequences	MOISST++Calibration + trajectory	22.26	0.794	0.376	26.44	0.859	0.242	0.948	8.54	73.30	0.100
PandaSetMean over 5 sequences	OriginalDataset poses	21.91	0.699	0.520	23.54	0.774	0.318	1.164	5.53	50.50	0.311
PandaSetMean over 5 sequences	MOISST++Calibration + trajectory	22.67	0.738	0.453	24.73	0.818	0.278	0.949	5.96	54.00	0.272
WaymoMean over 5 sequences	OriginalDataset poses	23.28	0.756	0.493	23.13	0.775	0.432	1.168	5.98	46.70	0.420
WaymoMean over 5 sequences	MOISST++Calibration + trajectory	24.40	0.781	0.447	25.02	0.809	0.367	1.065	6.46	69.50	0.357

Download scores CSV Download scores JSON Score definitions and aggregation

Evaluation protocol

Each dataset contributes five sequences. Dataset scores come from reference/current/*_pose.csv in the benchmark repository. Overall scores are the arithmetic mean of the unrounded twenty-sequence records for each pose set.

MOISST++ refines sensor extrinsics and ego trajectory; Original is generated locally from the downloaded dataset poses. Rendering evaluation uses Nerfacto with LiDAR depth supervision and Splatfacto with SfM + LiDAR initialization, with pose optimization disabled. PSNR, SSIM and LPIPS are part of the pose benchmark.

SfM triangulates with the supplied camera poses and intrinsics fixed. P2M precision uses a 0.15 m nearest-vertex threshold; surface distance uses triangle surfaces. These scores measure consistency rather than independent absolute pose ground truth.