DIMED / README.md
hljiang's picture
Upload DIMED dataset and per-image evaluation results
ca33145 verified
|
Raw History Blame Contribute Delete
8.13 kB
---
tags:
- depth
- intrinsics
- image
- geometry
- metric-depth
size_categories:
- 1K<n<10K
license: cc-by-4.0
pretty_name: DIMED
---
Processed versions of the open-source TUM RGB-D dataset, built for diagnosing **global-scale errors of single-image metric depth models**
(the model gets the shape right but the absolute scale wrong). They complement the seven metric benchmarks released with MoGe-2
([Ruicheng/monocular-geometry-evaluation](https://huggingface.co/datasets/Ruicheng/monocular-geometry-evaluation)) and use the same file layout,
so the MoGe evaluation data loader reads them unchanged. Evaluation code: <https://github.com/antigravity-tech/DIMED>. Project page: <https://antigravity-tech.github.io/DIMED/>.
This repository also holds the **per-image results** of all model evaluations in the paper (folder `results/`, see below).
| Dataset | Source | Publication | Num images | Storage Size | Note |
| --- | --- | --- | --- | --- | --- |
| TUM-Family | [TUM RGB-D](https://cvg.cit.tum.de/data/datasets/rgbd-dataset), Freiburg 3 structure / texture family | [1] | 514 | 72 MB | 8 sequences (`structure` / `nostructure` x `texture` / `notexture` x `near` / `far`). Frames kept when the camera moved >= 0.10 m or rotated >= 10 deg since the last kept frame. **A diagnostic about cue availability, not an everyday-scene benchmark; 8 sequences are too few to rank models.** |
| TUM-Ref | [TUM RGB-D](https://cvg.cit.tum.de/data/datasets/rgbd-dataset), `freiburg3_long_office_household` | [1] | 215 | 40 MB | Cue-rich office reference recorded with the same sensor, same frame rule. |
### TUM-Family: per sequence
| Sequence (folder) | Source sequence | Num images |
| --- | --- | --- |
| `fr3_nostruct_notex_far` | `freiburg3_nostructure_notexture_far` | 30 |
| `fr3_nostruct_notex_near` | `freiburg3_nostructure_notexture_near_withloop` | 116 |
| `fr3_nostruct_tex_far` | `freiburg3_nostructure_texture_far` | 42 |
| `fr3_nostruct_tex_near` | `freiburg3_nostructure_texture_near_withloop` | 131 |
| `fr3_struct_notex_far` | `freiburg3_structure_notexture_far` | 43 |
| `fr3_struct_notex_near` | `freiburg3_structure_notexture_near` | 42 |
| `fr3_struct_tex_far` | `freiburg3_structure_texture_far` | 57 |
| `fr3_struct_tex_near` | `freiburg3_structure_texture_near` | 53 |
TUM depth is short-range: for TUM-Family the 1st / 50th / 99th percentile is 0.74 / 1.45 / 4.40 m (maximum 8.0 m). In the two `nostruct_notex` sequences the image shows only a flat, plain surface, so the absolute
scale is not identifiable from a single image: read results there as "which prior did the model fall back to", not as failure.
Do not average the 8 sequences into one number without a sequence-level confidence interval.
## Format
Each sample is a directory `<dataset>/<sequence>/<id>/` containing
| File | Content |
| --- | --- |
| `image.jpg` | RGB image (640x480), already undistorted. |
| `depth.png` | Metric depth in **metres**, MoGe's per-image log-encoded 16-bit PNG: code 0 = no depth, `near` / `far` stored as PNG text chunks. Read it with `read_depth` from [moge/utils/io.py](https://github.com/microsoft/MoGe/blob/main/moge/utils/io.py). |
| `meta.json` | Normalised pinhole intrinsics `[[fx/W, 0, cx/W], [0, fy/H, cy/H], [0, 0, 1]]`. |
| `source.json` | Pointer to the original frame (RGB and depth file names inside the TUM sequence). |
Each dataset folder also has `.index.txt` (one sample per line) and `manifest.json` (selection rule, counts, intrinsics).
The depth maps are dense with holes (median 88% valid for TUM-Family, 84% for TUM-Ref).
## Notes
**TUM (TUM-Family, TUM-Ref).** Depth is the original `png / 5000` metres. Intrinsics are the official Freiburg 3 colour-camera values
(fx 535.4, fy 539.2, cx 320.1, cy 247.6); the Freiburg 3 images are distributed undistorted. Selection: RGB and depth paired by nearest timestamp
(|dt| <= 20 ms), ground-truth pose interpolated at the RGB time, frame kept by the motion rule above, >= 500 valid depth pixels. The selection uses poses only.
## Per-image results (`results/`)
| Path | Content |
| --- | --- |
| `results/main/<run>/<benchmark>.jsonl` | 36 model evaluations on the 7 MoGe-2 benchmarks (NYUv2, KITTI, ETH3D, iBims-1, DDAD, DIODE, HAMMER; 4,406 images) |
| `results/tum_family/<run>/TUM-Family.jsonl`, `results/tum_ref/<run>/TUM-Ref.jsonl` | the same models on the two TUM sets |
| `results/<set>/<run>/results.json` | per-benchmark means |
| `results/shared_failure_labels.csv` | for each benchmark image: `k` = number of the 11 architectures (with ground-truth intrinsics) whose metric δ1 < 0.5, tier (none / specific 1–2 / intermediate 3–5 / shared ≥ 6), and the 11 failure flags |
Each line of a `.jsonl` file (floats rounded to 6 significant digits):
| Field | Content |
| --- | --- |
| `filename` | image identifier inside the benchmark |
| `metric_scale` | the alignment scale `s*` = median of ground truth / prediction over valid pixels, so `s* · prediction ≈ ground truth`; `s* > 1` means depth is underestimated |
| `depth_metric` | `rel`, `delta1`, `delta2`, `delta3`, `sq_rel`, `rmse`, `rmse_log`, `log10` of the prediction as is (metric) |
| `depth_scale_invariant` | the same eight metrics after multiplying the prediction by `s*` (structure) |
| `scales` | five alignment scales, compared in Appendix C of the paper: `s_med` (= `metric_scale`), `s_lmean` = exp(−mean(ln p − ln g)) (log-MSE optimal), `s_rom` = median(g) / median(p) (ratio of medians), `s_moge64` and `s_full` = MoGe's solver `align_depth_scale` (minimises Σ\|s p − g\|/g) on the 64×64 mask-aware grid of the MoGe-2 protocol and on all valid pixels |
| `sa` | `rel`, `delta1` after multiplying the prediction by `s_lmean`, `s_rom`, `s_moge64` and `s_full` |
| `log_stats` | `log_mse`, `log_mean`, `log_var` of ln prediction − ln ground truth over the valid pixels with prediction > 0; `n` (these pixels), `n_bad_pred` (pixels with prediction ≤ 0) |
| `far10`, `n_pix` | share of valid pixels with prediction / ground truth > 10; number of valid pixels |
| `pair`, `s_bg`, `s_obj`, `n_bg_pixels`, `n_obj_pixels` | HAMMER only: the frame of the same view with / without objects, the five scales on the background / object pixels, and their pixel counts |
`results.json` in each run folder holds the per-benchmark means; the scale statistics are `bias` = mean(ln(1/s*)), `dispersion` = std(ln s*) and `abs_err` = mean(\|ln s*\|).
Run names: `<model>` without intrinsics, `<model>_withK` with ground-truth intrinsics (`metric3d_v2` and `metricanything_depthmap` always use them). Scripts that recompute the paper's numbers from these files are in `analysis/` of the code repository.
Download only the results: `hf download hljiang/DIMED --repo-type dataset --include "results/*" --local-dir .`
## Download
Make sure `huggingface_hub` is installed. If not installed:
```bash
pip install huggingface_hub
```
Download:
```bash
hf download hljiang/DIMED --repo-type dataset --local-dir /PATH/TO/DOWNLOAD/DIRECTORY
```
Extract the zip files (each dataset has a separate zip file; each must go into its own folder because the archives have no top-level directory):
```bash
cd /PATH/TO/DOWNLOAD/DIRECTORY
for n in TUM-Family TUM-Ref; do mkdir -p "$n" && unzip -q "$n.zip" -d "$n"; done
```
## Licence
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), inherited from TUM RGB-D ("Unless stated otherwise, all data in the TUM RGB-D benchmark is licensed under CC BY 4.0"). Please attribute TUM and check the original licence before reuse.
## Citation
```bibtex
@article{jiang2026averages,
title = {What the Averages Hide: Monocular Metric Depth Models Still Fail in Everyday Scenes, Mostly on Scale},
author = {Jiang, Hualie and Luo, Yujie and Song, Ziyang and Li, Ximeng and Lou, Zhiqiang and Xu, Rui},
journal = {arXiv preprint arXiv:2611.20518},
year = {2026}
}
```
## Reference
[1] Sturm, J., Engelhard, N., Endres, F., Burgard, W. and Cremers, D., 2012. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (pp. 573-580). IEEE.