README.md

---
license: cc-by-4.0
pretty_name: MetricBench
task_categories:
- depth-estimation
size_categories:
- 10K<n<100K
tags:
- benchmark
- visual-geometry
- metric-scale-recovery
- 3d-reconstruction
- camera-pose-estimation
- autonomous-driving
- embodied-ai
---

# MetricBench

[![Official Code](https://img.shields.io/badge/GitHub-Official_Code-181717?logo=github&logoColor=white)](https://github.com/yanglei18/UniMetric)
[![arXiv Paper](https://img.shields.io/badge/arXiv-2610.16945-B31B1B?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2610.16945)
[![ModelScope Dataset](https://img.shields.io/badge/ModelScope-MetricBench-624AFF)](https://www.modelscope.ai/datasets/yanglei_thu/MetricBench)

**A benchmark for metric-scale visual geometry across driving, roadside, and embodied scenarios.**

This is the **official MetricBench release** accompanying [UniMetric](https://arxiv.org/abs/2610.16945). MetricBench brings together **12 dataset subsets, 694 sequences, and 51,339 frames** to evaluate whether predicted depth, 3D structure, and camera motion are accurate in physical units.

## Associated paper

**[UniMetric: Universal Metric Scale Recovery for Visual Geometry Foundation Models](https://arxiv.org/abs/2610.16945)**

**Authors:** Lei Yang, Yuner Zhang, Ziyang Liu, Ziying Song, Kai Wu, Yiyang Gao, Jiangtao Su, and Yiming Li.  
**Publication:** Technical report, 2026 · [arXiv:2610.16945](https://arxiv.org/abs/2610.16945).

UniMetric is a training-free framework that combines uncertain object-size priors from a vision-language model with reconstructed geometry to recover a shared metric scale for visual geometry foundation models. MetricBench is the accompanying benchmark for evaluating metric depth, 3D reconstruction, and camera trajectories.


## Why MetricBench?

- **Metric accuracy across three geometric tasks.** Evaluate depth estimation, 3D reconstruction, and camera pose against sensor-derived references, without fitting predicted scale to ground truth.
- **Diverse viewpoints and scene scales.** Cover moving vehicles, static roadside cameras, drones, quadrupeds, humanoids, and manipulation settings in one evaluation suite.
- **Prepared data and reusable semantic priors.** Use a consistent RGB / point-cloud / metadata layout together with the 51,339 precomputed VLM responses bundled in the official code repository.

## Dataset composition

| Archive | Dataset folder | Sequences | Frames |
|:--|:--|--:|--:|
| `autonomous_driving.zip` | `Argoverse` | 70 | 6,613 |
| | `dair-v2x-i` | 86 | 3,704 |
| | `KITTI` | 96 | 5,211 |
| | `nuScenes` | 116 | 4,627 |
| | `v2x-radar-i` | 30 | 1,092 |
| | `Waymo` | 149 | 5,951 |
| `embodied_ai.zip` | `3EED` | 20 | 3,844 |
| | `CEAR` | 20 | 5,670 |
| | `DROID` | 45 | 3,102 |
| | `FusionPortable` | 5 | 3,847 |
| | `HumanoidEveryday` | 30 | 3,949 |
| | `JRDB` | 27 | 3,729 |
| **Total** | **12 subsets** | **694** | **51,339** |

These are the prepared MetricBench evaluation subsets, **not the complete official splits of the source datasets**. In particular, 3EED contains 12 drone and 8 quadruped sequences; HumanoidEveryday contains 30 selected G1 episodes. The two static roadside subsets, `dair-v2x-i` and `v2x-radar-i`, are evaluated for depth and reconstruction and do not participate in ego-pose evaluation. See the paper for the sampling and evaluation protocols.

## Download and data layout

Download both archives for the complete 12-subset release:

- [autonomous_driving.zip](https://huggingface.co/datasets/yanglei18/MetricBench/resolve/main/autonomous_driving.zip) — approximately **50.2 GB**, 6 subsets.
- [embodied_ai.zip](https://huggingface.co/datasets/yanglei18/MetricBench/resolve/main/embodied_ai.zip) — approximately **7.1 GB**, 6 subsets.

After following the [official environment setup](https://github.com/yanglei18/UniMetric#installation), run the following **from the UniMetric repository root**:

```bash
hf download yanglei18/MetricBench \
  autonomous_driving.zip embodied_ai.zip \
  --repo-type dataset --local-dir ./data/metric_bench

for group in autonomous_driving embodied_ai; do
  unzip "data/metric_bench/${group}.zip" -d data/metric_bench
  mv "data/metric_bench/${group}"/*/ data/metric_bench/
  rmdir "data/metric_bench/${group}"
done
```

The extraction step removes the archive-level group folders. The default code expects dataset folders directly under `data/metric_bench/`:

```text
data/metric_bench/
├── Argoverse/
├── dair-v2x-i/
├── KITTI/
├── nuScenes/
│   └── <sequence>/
│       ├── <frame>_camera1.jpg
│       ├── <frame>.pcd
│       └── <frame>.yaml
├── v2x-radar-i/
├── Waymo/
├── 3EED/
├── CEAR/
├── DROID/
├── FusionPortable/
├── HumanoidEveryday/
└── JRDB/
```

Each prepared frame has an RGB image, a corresponding point cloud, and YAML metadata used by the dataset loaders for calibration and pose information. Preserve the exact folder names and capitalization. The default layout requires **no dataset-path edits** in the supplied configurations.

## Run with precomputed VLM responses

The [official UniMetric repository](https://github.com/yanglei18/UniMetric) includes [`precomputed/vlm_results_12datasets.zip`](https://github.com/yanglei18/UniMetric/blob/main/precomputed/vlm_results_12datasets.zip), containing **51,339 precomputed VLM responses covering all 12 released subsets**. This cache is distributed with the code repository, separately from the two data archives above.

With the default cache-reuse setting, the runner automatically restores and reuses matching responses. **No VLM API key is needed for the covered release frames.** Geometry inference, segmentation, scale recovery, and evaluation still run locally.

After installing the environment and downloading the data:

```bash
conda activate unimetric
python get_metric_scale.py
```

This default command runs **LingBot-Map on nuScenes**, using `configs/nuscenes/lingbot_map_long.yaml`; it does not launch all 12 subsets at once. Results are saved under `outputs/nuScenes/`. Use the supplied dataset/model configurations to evaluate the other subsets and GFMs; see the [code README](https://github.com/yanglei18/UniMetric#readme) for the full installation and execution guide.

New images without cached responses, or deliberate regeneration of VLM outputs, require a configured VLM API. Model downloads and any upstream model-access requirements are separate from VLM API access.

## Evaluation protocol

MetricBench measures geometry at its predicted physical scale. Depth and reconstruction are evaluated without ground-truth scale fitting; camera trajectories use **SE(3) alignment**, which preserves scale. Report the selected subset, model, and configuration when comparing results. Use the official evaluation pipeline to keep preprocessing, reference construction, and aggregation consistent.

## Source datasets and usage

MetricBench is assembled from the source datasets listed above. Their original licenses, attribution requirements, and usage conditions continue to apply to the underlying data. The repository license metadata does not replace the terms of those source datasets. Please cite the original datasets you use as well as UniMetric.

## Citation

```bibtex
@article{yang2026unimetric,
  title   = {UniMetric: Universal Metric Scale Recovery for Visual Geometry Foundation Models},
  author  = {Yang, Lei and Zhang, Yuner and Liu, Ziyang and Song, Ziying and Wu, Kai and Gao, Yiyang and Su, Jiangtao and Li, Yiming},
  journal = {arXiv preprint arXiv:2610.16945},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.16945}
}
```

For questions about the prepared data or evaluation, please open an issue in the [official code repository](https://github.com/yanglei18/UniMetric/issues).