Clarify depth PNG placement in dataset card

README.md

@@ -1,14 +1,24 @@
 ---
+pretty_name: IH-Depth
 license: cc-by-4.0
 task_categories:
-- depth-estimation
+  - depth-estimation
+size_categories:
+  - n<1K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: scenes_train.csv
+      - split: test
+        path: scenes_test.csv
 ---
 
 # Dataset Card for IH-Depth
 
 IH-Depth is a benchmark for metric depth estimation from long-wave hyperspectral imagery (LWHSI) in off-road scenes. The benchmark is derived from the [Invisible Headlights (IH) dataset](https://registry.opendata.aws/darpa-invisible-headlights/) and pairs raw LWHSI scenes with sparse LiDAR-projected depth labels, cylindrical camera geometry, and point/image correspondence files.
 
-IH-Depth contains 51 released scenes split into 41 training scenes and 10 test scenes. The benchmark is distributed as an overlay on top of the raw IH dataset: after unpacking IH-Depth into the raw IH root, each benchmark scene adds depth and geometry artifacts beside the matching raw `.hdr` and `.bsq` LWHSI files.
+IH-Depth contains 51 released scenes split into 41 training scenes and 10 test scenes. Download the dataset into the same directory as the raw IH dataset so each released depth PNG lands beside the matching raw `.hdr` and `.bsq` LWHSI files.
 
 More details on usage, evaluation, and reference baselines can be found in the [IH-Depth GitHub repository](https://github.com/cvail-research/ih-depth-dataset).
 
@@ -22,7 +32,7 @@ Each released scene contains three IH-Depth artifacts named with the same stem a
 <raw_lwhsi_stem>_corresp.txt
 ```
 
-The depth maps are sparse metric labels projected from LiDAR into the LWHSI image. Depth PNGs use a KITTI-style 16-bit encoding:
+The depth maps are sparse metric labels projected from LiDAR into the LWHSI image. Depth PNGs follow a KITTI-style 16-bit encoding:
 
 ```text
 dtype: uint16
@@ -31,11 +41,11 @@ stored_value = round(128 * depth_m)
 depth_m = stored_value / 128
 ```
 
-The `.cyl` files provide per-scene cylindrical camera geometry. The `_corresp.txt` files provide point/image correspondences used to support the released geometry and labels.
+The `.cyl` files provide per-scene cylindrical camera geometry. The `_corresp.txt` files provide point/image correspondences used to support the released geometry and labels. More details about these sidecar formats can be found in the [IH-Depth GitHub repository](https://github.com/cvail-research/ih-depth-dataset/ihd/example).
 
 ## Dataset Structure
 
-IH-Depth follows the raw IH folder structure. A released scene has the form:
+Only the depth PNGs are meant to be stored in the raw IH tree beside the matching raw files. A released depth-labeled scene has the form:
 
 ```text
 <RAW_IH_ROOT>/
@@ -45,8 +55,6 @@ IH-Depth follows the raw IH folder structure. A released scene has the form:
         <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 contains:
@@ -61,6 +69,8 @@ README.md
 
 The scene CSV files define the IH-Depth benchmark splits and artifact paths. These split files are part of IH-Depth and are not part of the raw IH dataset.
 
+The benchmark `.cyl` and `_corresp.txt` files should be kept separate from the raw IH root. Raw IH already contains original `.cyl` and `_corresp.txt` files for some scenes, so placing benchmark geometry and correspondence files into the same namespace would be ambiguous.
+
 ## Data Stats
 
 To characterize scene diversity, RGB visualizations of the released scenes were assigned to six primary scene categories and manually reviewed for consistency: