TLDR
- NVIDIA posted MetricTrail 1.0 to Hugging Face on November 5, 2026.
- The tool estimates camera paths and depth from RGB image sequences.
- It is licensed for noncommercial research and evaluation only.
- The model weights are not yet in the repository and will come in a separate release.
- On KITTI, calibrated stereo with loop correction scored 3.00 m average error.
NVIDIA has uploaded a new research model called MetricTrail to Hugging Face. The release date listed on the model card is November 5, 2026. The company says the model was developed by NVIDIA.
MetricTrail estimates how a camera moves and how far away objects are in each frame. It works from long sequences of ordinary color images. Researchers can use it for visual odometry and 3D reconstruction work.
The system combines three parts. These are single-image metric depth, learned sparse correspondences, and local geometric optimization. It can use camera settings that a user supplies, or it can estimate them from the images. It can also take rectified stereo pairs.
What Is in the Release
The repository offers a command-line tool. It exports camera trajectories and saves per-frame depth data. It does not export dense point clouds or meshes.
The model weights are not included in the source repository. The checkpoint, called DINO-DPVO, is planned for a separate Hugging Face release. It holds about 3.43 million learned parameters.
The model builds on DPVO, a visual odometry method. It adds a frozen DINOv3 image encoder and a frozen MoGe-2 depth model. MoGe-2 predicts metric depth for each image.
NVIDIA trained the system for 15,000 steps on 16 A100 GPUs. Training took about seven hours. The data came from five synthetic or game-based sources, including TartanAir, TartanGround, CarlaOcc, LightwheelOcc, and OmniWorld-Game.
Use of the weights falls under the NVIDIA License. It limits use to noncommercial research or evaluation. The DPVO base model uses the MIT License, and some NVIDIA-written code uses Apache 2.0.
Setup and Requirements
The pipeline needs a CUDA device. The documented setup is Linux x86_64 with Python 3.10, PyTorch 2.5.1, and CUDA 12.4. NVIDIA lists the Ampere A100 as the supported hardware.
Users must build several DPVO CUDA extensions from source before running it. No precompiled versions are included. Other GPUs and minimum memory needs have not been established.
The input must be at least 16 images in order, with a positive frame rate. The tool reads PNG frames or KITTI NPZ files. It does not decode video files directly.
An optional loop correction step runs separately on the CPU. It uses candidate frame pairs supplied by the user. NVIDIA says this newer step has not been checked against benchmarks.
Benchmark Results
NVIDIA tested the model on several public datasets. Trajectory error is measured after aligning each path to the reference. Lower numbers mean smaller errors.
On KITTI Odometry, the RGB-only setup scored 14.05 m average error without loop correction. With loop correction it scored 9.78 m. Supplied camera settings brought it to 7.39 m, and calibrated stereo reached 3.00 m.
On the VBR dataset, a focal length refinement step cut error from 33.51 m to 16.36 m without loop correction. With loop correction it dropped from 27.95 m to 10.88 m. The refinement averaged 7.1 minutes per sequence.
Indoors, the results were measured in centimeters. On TUM RGB-D with RGB-only calibration, error fell from 8.4 cm locally to 4.4 cm with global bundle adjustment and refinement. On Wild-SLAM MoCap, the scores were 2.70 cm and 1.65 cm.
For dense reconstruction, NVIDIA reported F1 scores of 97.60 percent on 7Scenes and 95.12 percent on Oxford Spires. TUM-Dynamic scored 89.58 percent. NVIDIA says its TUM-Dynamic scores trail leading methods.
On KITTI video depth, the absolute relative error was 0.057. The accuracy at the 1.25 ratio threshold was 97.8 percent.
On KITTI-02, the model ran at 12.70 frames per second. With supplied camera settings, the first output came in about 0.7 seconds. With RGB-only calibration, it took about 175 seconds.
NVIDIA lists known limits. Low parallax, fast rotation, rolling shutter, and motion blur can hurt results. The model has no explicit mask for moving objects.
NVIDIA says the model does not guarantee correct geometry. Metric scale remains an estimate. The company says users should test it on their own data before any wider use.
