N
NVIDIA / nvidia/MetricTrail · dino_dpvo.pth@dad495a8f2f6191500209f3a0bd395620c7e301d

MetricTrail 1.0

NVIDIA research checkpoint released on November 5 to estimate camera motion and metric depth. It contains about 3.43 million learned parameters, excluding external DINOv3 and MoGe-2 models. Although the card still announces a future download, the dino_dpvo.pth artifact is already published with the documented hash.

01 / SUMMARY

The essential

ORIENTATIONVisual odometry and metric depth

Primary use stated or inferred cautiously from official documentation.

CONTEXTTime-ordered inventory of at least 16 images

Maximum input capacity when published by the source.

EXITNPZ files and JSON records

Documented maximum generation limit.

ACCESSHugging Face · downloadable checkpoint

Verified availability channels.

MODEL TYPEVision and geometry · Robotics

Primary functional family and verified specialties.

LICENSENVIDIA OneWay Noncommercial License · dependencies have their own licenses

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDnvidia/MetricTrail · dino_dpvo.pth@dad495a8f2f6191500209f3a0bd395620c7e301d
Model typeVision and geometry · Robotics
Access modelOpen weights
LicenseNVIDIA OneWay Noncommercial License · dependencies have their own licenses
DeploymentSelf-hosted
Release05 NOV 2026
Knowledge cutoffNot published
EntranceRGB image sequences · Camera parameters
ExitCamera trajectories · Depth maps
Context windowTime-ordered inventory of at least 16 images
maximum outputNPZ files and JSON records
ReasoningNot published
Published tools
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsHugging Face · PyTorch
BEST SUITED FOR

Odometry and geometric reconstruction research with ordered RGB sequences. Configuration accepts known or estimated calibration and optional rectified stereo. The CLI exports trajectories and depth caches; it does not directly export a mesh or dense point cloud.

WORTH MONITORING

Weights are restricted to noncommercial research or evaluation, and dependencies retain their own licenses. The documented environment uses Linux and CUDA with extensions that must be compiled. The new loop correction from external candidates has no demonstrated quantitative equivalence to the benchmark workflow. It does not establish safe robot navigation.

03 / CAPABILITIES

What it can do

01

Orientation

Odometry and geometric reconstruction research with ordered RGB sequences. Configuration accepts known or estimated calibration and optional rectified stereo. The CLI exports trajectories and depth caches; it does not directly export a mesh or dense point cloud.

02

Context

Time-ordered inventory of at least 16 images · NPZ files and JSON records

03

Tools and integration

Not published

04

Access

Hugging Face · downloadable checkpoint

Input modalities
RGB image sequencesCamera parameters
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputNo published official API rateStandard API rate
Cached inputNot publishedReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputSelf-hosted infrastructure costMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

The weights require your own infrastructure. Cost depends on memory, precision, sequence length and load. Review the license and runtime code before deployment. A published performance figure does not guarantee the same cost or speed on another machine.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The evaluation reports 14.05 m mean ATE on KITTI Odometry with RGB-only calibration and no loop closure. It measures RMSE after one Sim(3) alignment per sequence, not absolute accuracy without alignment. Results with other loop closures, stereo or dense reconstruction use different protocols.

BenchmarkResultMetricSource
KITTI Odometry ATE (RGB-only, no LC)14,05 mError; lower is betterView source ↗

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

MetricTrail 1.0

Version added to Inferama's verified catalog.

07 / SOURCES

Traceability