Add RAI fields to Croissant metadata

Metadata-only update: add Croissant RAI fields and provenance activity. Dataset files, manifests, and benchmark contents are unchanged.

ih-depth.croissant.json

@@ -47,7 +47,9 @@
     "separator": "cr:separator",
     "source": "cr:source",
     "subField": "cr:subField",
-    "transform": "cr:transform"
+    "transform": "cr:transform",
+    "prov": "http://www.w3.org/ns/prov#",
+    "dataBiases": "cr:dataBiases"
   },
   "@type": "sc:Dataset",
   "name": "IH-Depth",
@@ -292,7 +294,31 @@
   "rai:dataLimitations": [
     "IH-Depth is intended for research on monocular metric depth estimation in off-road longwave infrared hyperspectral imagery. It should not be treated as a general-purpose depth benchmark for on-road RGB, indoor, aerial, or consumer-camera settings.",
     "The dataset contains sparse LiDAR-projected depth labels rather than dense ground-truth depth. Metrics should therefore be interpreted at valid projected LiDAR pixels only.",
-    "The dataset covers a limited set of physical locations, sensors, seasons, and environmental conditions inherited from the source IH data and from the 51 included IH-Depth scenes."
+    "The dataset covers a limited set of physical locations, sensors, seasons, and environmental conditions inherited from the source IH data and from the 51 included IH-Depth scenes. Models trained on this dataset may not generalize to unseen geography, sensor calibration, weather, vegetation, terrain, or vehicle/sensor mounting configurations.",
+    "Depth labels depend on per-scene cylindrical registration, automatic filtering, cleanup, and qualitative validation. Residual calibration errors, occlusion artifacts, LiDAR noise, or projection errors may remain.",
+    "The dataset should not be used as the sole basis for safety-critical autonomous navigation, obstacle avoidance, or deployment decisions without independent validation on the target platform and environment."
   ],
-  "rai:personalSensitiveInformation": "IH-Depth does not intentionally include personal or sensitive information. The data consists of off-road LWHSI imagery, LiDAR-derived depth labels, camera geometry, correspondences, manifests, and related source or diagnostic files."
+  "rai:personalSensitiveInformation": "IH-Depth does not intentionally include personal or sensitive information. The data consists of off-road LWHSI imagery, LiDAR-derived depth labels, camera geometry, correspondences, manifests, and related source or diagnostic files. The source scenes may contain vehicles, objects, infrastructure, or outdoor locations, but the dataset is not designed to identify people and no demographic, biometric, medical, financial, political, religious, or other personal attributes are provided.",
+  "rai:dataBiases": [
+    "The data is biased toward off-road scenes represented in the Invisible Headlights collection and in the 51 included IH-Depth scenes. Terrain, vegetation, structures, weather, season, time of day, and geography are not uniformly sampled.",
+    "The source IH dataset used two different LWHSI sensors with different spectral ranges and spatial resolutions, so performance may be sensor-dependent.",
+    "The valid depth labels are sparse and LiDAR-projection-dependent. Surfaces or regions poorly observed by LiDAR, affected by occlusion, or removed by filtering may be underrepresented in evaluation.",
+    "Scene inclusion reflects registration quality and qualitative validation decisions, which may bias the benchmark toward scenes where LiDAR-to-image alignment is visually reliable."
+  ],
+  "rai:dataUseCases": [
+    "Benchmarking supervised monocular metric depth estimation from longwave infrared hyperspectral imagery under challenging off-road sensing conditions.",
+    "Studying depth estimation under sparse LiDAR-projected supervision and evaluating models at valid projected LiDAR pixels.",
+    "Research on dataset construction, camera/LiDAR registration, cylindrical camera geometry, and label-quality filtering for hyperspectral thermal depth datasets.",
+    "Exploring future hyperspectral learning approaches that leverage spectral features for depth estimation.",
+    "Not recommended for direct safety-critical deployment, human identification, surveillance, demographic inference, or claims of general performance outside the documented off-road LWHSI setting."
+  ],
+  "rai:dataSocialImpact": "Potential positive impacts include improving reproducibility and enabling research on passive sensing for off-road robotics, especially in conditions where RGB or active sensing may be insufficient. Potential negative impacts include overclaiming model robustness, using the benchmark as evidence for real-world autonomous navigation safety without deployment-specific validation, or adapting the data for surveillance or military applications. The dataset should be used with explicit reporting of its sparse-label nature, off-road scope, registration pipeline, and evaluation protocol.",
+  "rai:hasSyntheticData": false,
+  "prov:wasGeneratedBy": [
+    {
+      "@type": "prov:Activity",
+      "name": "IH-Depth label generation",
+      "description": "High-resolution LiDAR was registered and projected into LWHSI image space using per-scene cylindrical camera geometry; inclusion decisions were frozen after automatic filtering, cleanup, and qualitative validation."
+    }
+  ]
 }