IH-Depth

IH-Depth v0 is a curated LWIR/LWHSI-LiDAR benchmark derived from the Invisible Headlights dataset. It contains 51 included off-road scenes with sparse LiDAR-projected metric depth labels, per-scene cylindrical camera geometry, correspondence files, and release manifests.

Repository Status

• Hugging Face repository: SemilleroCV/ih-depth
• Version: v0 (0.0.0 for Croissant semver metadata)
• Published: 2026-05-07
• License: CC BY 4.0
• Total scenes in frozen manifest: 306
• Included scenes: 51
• Deferred scenes: 202
• Excluded scenes: 53
• Train scenes: 41
• Test scenes: 10
• Canonical included-scene triples: 51

Files

Top-level metadata files:

• scenes_manifest.csv: one row per included scene, with repository-relative paths to the canonical files.
• scenes_train.csv: train split scene manifest used by the Hugging Face dataset viewer.
• scenes_test.csv: test split scene manifest used by the Hugging Face dataset viewer.
• frozen_manifest_included_51.csv: full frozen-manifest rows for the 51 included scenes.
• 07_frozen_manifest_v0.csv: frozen release manifest for all 306 candidate scenes.
• release_summary.json: machine-readable release counts and metadata-correction scope.
• ih-depth.croissant.json: Croissant metadata describing this repository layout and RAI metadata.

Canonical per-scene files:

• depth_label.npz: sparse metric depth labels projected from LiDAR into LWHSI image coordinates.
• correspondences.txt: image-to-geometry correspondences used by the registration pipeline.
• camera.cyl: cylindrical camera geometry for the scene.

Some scenes also include source or diagnostic assets inherited from release preparation, such as .bsq, .hdr, .las, .png, .npy, *_label.npz, and *_all.cyl files. The canonical benchmark interface is the 51 rows in scenes_manifest.csv.

Minimal Usage

from huggingface_hub import hf_hub_download
import pandas as pd
import numpy as np

repo_id = "SemilleroCV/ih-depth"
manifest_path = hf_hub_download(repo_id, "scenes_manifest.csv", repo_type="dataset")
scenes = pd.read_csv(manifest_path)
row = scenes.iloc[0]

depth_path = hf_hub_download(repo_id, row.depth_label_path, repo_type="dataset")
depth_npz = np.load(depth_path)
print(depth_npz.files)

The depth_label.npz internal array names should be inspected with np.load(...).files, because downstream code should not assume undocumented internal key names.

Evaluation Scope

IH-Depth provides sparse LiDAR-projected depth labels, not dense ground-truth depth. Metrics should be computed only at valid projected LiDAR pixels. The dataset is intended for research on monocular metric depth estimation in off-road longwave infrared hyperspectral imagery and should not be treated as a general-purpose RGB, indoor, on-road, aerial, or safety-certification benchmark.

Provenance

IH-Depth is derived from the Invisible Headlights dataset and from the registration, filtering, cleanup, and validation pipeline documented by the release manifests. Scene inclusion was frozen in 07_frozen_manifest_v0.csv and the benchmark train/test split was frozen in scenes_train.csv and scenes_test.csv.

Responsible Use

The dataset should not be used as the sole basis for safety-critical autonomous navigation, obstacle avoidance, or deployment decisions. Report the sparse-label nature, off-road scope, registration process, and evaluation protocol when using the dataset.