Datasets:

SemilleroCV
/

ih-depth

IH-Depth is a curated LWIR/LWHSI-LiDAR benchmark derived from the Invisible Headlights (IH) dataset. The released benchmark contains 51 off-road scenes, split into 41 training scenes and 10 test scenes. The dataset is released under the CC BY 4.0 license. This repository is the benchmark devkit and code for running the reference baseline methods for that release.

Getting Started

Download the dataset

First download and unpack the raw Invisible Headlights (IH) dataset. Then download IH-Depth and unpack it into the same root directory.

IH-Depth is a KITTI-style overlay on top of the raw IH tree. The benchmark depth maps, cylindrical camera files, and correspondence files follow the raw IH folder structure, so each IH-Depth scene artifact lands beside the corresponding raw LWHSI .hdr/.bsq files:

RAW_IH_ROOT/
  <collection>/
    <PathXX_DistStA>/
      <PathXX_StepYY_DistStA>/
        <raw_lwhsi_stem>.hdr
        <raw_lwhsi_stem>.bsq
        <raw_lwhsi_stem>_depth.png
        <raw_lwhsi_stem>.cyl
        <raw_lwhsi_stem>_corresp.txt

The release root also contains scenes_train.csv, scenes_test.csv, scenes_manifest.csv, release_summary.json, and this README. The train/test split CSVs define the IH-Depth benchmark splits; they are not part of the raw IH dataset.

IH-Depth train and test

For IH-Depth training and testing sets we release sparse LiDAR-projected metric depth labels, per-scene cylindrical camera geometry, and correspondence files. Each released scene has exactly these public benchmark artifacts, all sharing the original raw LWHSI stem:

<raw_lwhsi_stem>_depth.png
<raw_lwhsi_stem>.cyl
<raw_lwhsi_stem>_corresp.txt

Depth PNGs follow KITTI-style encoding contract with a small modification:

dtype: uint16
stored_value = round(128 * depth_m)
0 = invalid / unlabeled
depth_m = stored_value / 128

This x128 scale preserves the released maximum observed depth range (<415 m) while remaining KITTI-style 16-bit PNG encoding format. ihd/example/ contains the first-training-scene artifact bundle: a benchmark-format depth PNG plus the matching .cyl and correspondence .txt example files. The example is for structure and inspection.

Test Evaluation

We provide an evaluator to compute common depth estimation metrics (AbsRel, RMSE, $\delta$ and more) on the IH-Depth test set. First set up the benchmark environment:

uv sync

After unpacking IH-Depth over the raw IH root, create a compact ground-truth evaluation folder from scenes_test.csv:

uv run python -m ihd.utils.prepare_eval_split RAW_IH_ROOT GT_DIR --split_csv scenes_test.csv

Use --symlink to link depth PNGs instead of copying them. The helper only prepares depth PNGs because the evaluator does not consume .cyl or _corresp.txt files.

Run your model on the raw .hdr/.bsq input files and save prediction PNGs into a mirrored prediction tree:

GT_DIR/                                           # ground-truth root folder
├── <collection>/
│   ├── <PathXX_DistStA>/
│   │   ├── <PathXX_StepYY_DistStA>/
│   │   │   ├── <raw_lwhsi_stem>_depth.png       # ground-truth depth PNG
│   │   │   └── ...
│   │   └── ...
│   └── ...
└── ...

PREDICTION_DIR/                                   # prediction root folder
├── <collection>/
│   ├── <PathXX_DistStA>/
│   │   ├── <PathXX_StepYY_DistStA>/
│   │   │   ├── <raw_lwhsi_stem>_depth.png       # prediction depth PNG
│   │   │   └── ...
│   │   └── ...
│   └── ...
└── ...

Then run the evaluator:

uv run python ihd/ihd_evaluator.py GT_DIR PREDICTION_DIR --output_name stats_ihd.txt

Here, <raw_lwhsi_stem> means the original LWHSI filename stem for that scene. For example, if the original LWHSI file is named:

IHTest_202104_Path15_Step11_LWHSI1_collect0_DistStA.hdr

then the evaluated depth PNG must be named:

IHTest_202104_Path15_Step11_LWHSI1_collect0_DistStA_depth.png

The evaluator writes stats_ihd.txt inside PREDICTION_DIR, prints the aggregate summary to stdout, and emits:

PREDICTION_DIR/
  stats_ihd.txt
  errors_out/
  errors_img/
  depth_gt/
  depth_pred/
  input_preview/

If you also want to run the retained Depth Anything V2 baseline example below, install its extra dependencies with:

uv sync --extra depthanythingv2

Then, make inference with the following command:

uv run python -m baselines.learning_broadband.depthanythingv2 \
  --hdr /path/to/scene/IHTest_202104_Path15_Step11_LWHSI1_collect0_DistStA.hdr \
  --out-dir /tmp/ihd_predictions \
  --device cuda

That command writes a public-contract prediction PNG named <raw_lwhsi_stem>_depth.png that can be dropped into PREDICTION_DIR.

First-user flow

Download and unpack the raw IH dataset.
Download IH-Depth and unpack it into the raw IH root.
Use scenes_test.csv and ihd.utils.prepare_eval_split to prepare GT_DIR.
Run a model using the raw .hdr/.bsq input files.
Save predictions as <raw_lwhsi_stem>_depth.png in the mirrored PREDICTION_DIR.
Run uv run python ihd/ihd_evaluator.py GT_DIR PREDICTION_DIR.

Acknowledgement

We sincerely thank the authors of Concurrent Band Selection and Traversability Estimation From Long-Wave Hyperspectral Imagery in Off-Road Settings for releasing the IH dataset. We are grateful to the authors of Ozone-Cues-Mitigate-Reflected-Downwelling-Radiance-in-LWIR-Absorption-Based-Ranging, Depth Pro, UniK3D, UniDepth, Depth Anything V2, and KITTI-devkit for open-sourcing their code and models.

License

The code in this repository is released under the MIT license in LICENSE. The IH-Depth benchmark data released here is released under CC BY 4.0.