Update README.md

README.md

@@ -1,6 +1,8 @@
 ---
 pretty_name: HighBuild-1M
 license: other
+language:
+- en
 task_categories:
 - image-segmentation
 - object-detection
@@ -9,10 +11,14 @@ tags:
 - geospatial
 - building-height
 - aerial-imagery
+- overhead-imagery
 - earth-observation
 - single-view-height-estimation
 - building-height-estimation
 - spatial-generalization
+- building-segmentation
+- coco
+- tiff
 size_categories:
 - 10K<n<100K
 configs:
@@ -28,263 +34,52 @@ configs:
 
 # HighBuild-1M
 
-HighBuild-1M is a multi-continental high-resolution benchmark dataset for single-view building height estimation from overhead imagery. Each sample contains a 1024×1024 RGB overhead image, a spatially aligned float32 building-height mask, and COCO-style building instance annotations.
+HighBuild-1M is a multi-continental high-resolution benchmark dataset for single-view building height estimation from overhead imagery. Each sample contains a 1024 x 1024 RGB overhead image, a spatially aligned float32 building-height mask, and COCO-style building instance annotations.
 
 ## Dataset Statistics
 
-| Version | 1024×1024 tiles | Building instances | City groups | Countries/regions | Continents |
+| Version | 1024 x 1024 tiles | Building instances | City groups | Countries/regions | Continents |
 |---|---:|---:|---:|---:|---:|
-| HighBuild-1M | 70,266 | 6,050,823 | 26 | 12 | 6 |
+| HighBuild-1M full benchmark | 70,266 | 6,050,823 | 26 | 12 | 6 |
 
-The released benchmark covers 6 continents, 12 countries or regions, and 26 city groups, with 70,266 paired 1024×1024 tiles and 6,050,823 building instances.
+The full benchmark covers 6 continents, 12 countries or regions, and 26 city groups, with 70,266 paired 1024 x 1024 tiles and 6,050,823 building instances.
+
+## Reviewer Small Sample
+
+A small reviewer-inspection subset is available at:
+
+https://huggingface.co/datasets/feifei140729/small-sample
+
+The small sample follows the same triplet structure as the full dataset, including RGB images, float32 TIFF building-height masks, and COCO-style JSON annotations. It is intended for quick inspection of data quality, file organization, spatial alignment, and annotation format.
+
+## Dataset Viewer Note
+
+The Hugging Face Dataset Viewer may fail to render this dataset because HighBuild-1M is distributed as large WebDataset TAR shards containing JPEG images, float32 TIFF masks, and COCO-style JSON annotations rather than a single tabular dataset. This does not affect downloading, streaming, or inspecting the dataset. Please use the WebDataset loading instructions and manifest files below.
 
 ## Tasks
 
 HighBuild-1M supports the following tasks:
 
 1. **Single-view building height estimation**  
-   Input: one 1024×1024 RGB overhead image tile.  
+   Input: one 1024 x 1024 RGB overhead image tile.  
    Output: a spatially aligned float32 building-height map.
 
 2. **Building-wise height evaluation**  
    Predicted height maps can be aggregated within COCO-style building polygons to compute building-level MAE/RMSE.
 
-3. **Building segmentation / instance-level building understanding**  
-   COCO-style building polygons and bounding boxes can be used for semantic or instance-level building analysis.
+3. **Building segmentation and instance-level building understanding**  
+   COCO-style building polygons and bounding boxes can be used for semantic building segmentation, instance-level building analysis, and joint height-segmentation modelling.
 
 4. **Spatial generalization benchmarking**  
    The benchmark supports same-city, cross-city within-country, and cross-country evaluation protocols.
 
-   
 ## WebDataset and Compression
 
-For the Hugging Face hosted release, the recommended distribution format is
-WebDataset TAR shards generated from this staging directory:
+For the Hugging Face hosted release, the recommended distribution format is WebDataset TAR shards.
 
-```bash
-python3 scripts/build_hf_webdataset.py \
-  --input hf_upload_staging \
-  --output hf_upload_webdataset \
-  --clean \
-  --workers 8 \
-  --max-shard-gb 1.0
-```
-
-The conversion writes split-aware shards under:
+The hosted WebDataset release uses split-aware shards under:
 
 ```text
 data/webdataset/train/*.tar
 data/webdataset/validation/*.tar
-data/webdataset/test/*.tar
-```
-
-The conversion also writes
-`benchmark_v1/manifest_webdataset_tiles_1024.csv`, which maps each original
-tile id to its split, shard path, WebDataset key, and member filenames.
-
-Each WebDataset sample contains three members with the same key:
-
-```text
-<Continent>_<Country>_<City>__<tile_id>.jpg
-<Continent>_<Country>_<City>__<tile_id>.tiff
-<Continent>_<Country>_<City>__<tile_id>.json
-```
-
-The TAR shards themselves are not gzip/zstd-compressed, which keeps them
-streamable. Before each mask is written into a shard, the mask TIFF is
-losslessly recompressed with internal TIFF ZSTD compression:
-
-```text
-COMPRESS=ZSTD
-ZSTD_LEVEL=9
-PREDICTOR=3
-TILED=YES
-BLOCKXSIZE=512
-BLOCKYSIZE=512
-BIGTIFF=IF_SAFER
-```
-
-This changes only the TIFF storage encoding. No resampling, quantization, dtype
-conversion, or lossy compression is applied to mask values. The RGB JPEG images
-and COCO-style JSON annotations are copied into the shards without
-recompression.
-
-In the local staging package, mask TIFF files account for about 240 GB of the
-270 GB total. A local random sample of 50 masks compressed from 4.2 MB per mask
-to about 27 KB on average with identical GDAL checksums. Based on that sample,
-the mask portion is expected to shrink to roughly 1.5 GB, and the final
-WebDataset release is expected to be dominated by the existing JPEG imagery and
-JSON annotations.
-
-## Dataset Structure
-
-The original unsharded layout represented each sample by three matched files with the same tile basename:
-
-```text
-data/images/<Continent>_<Country>_<City>/<tile_id>.jpg
-data/masks/<Continent>_<Country>_<City>/masks/<tile_id>.tiff
-data/annotations/coco_json/<Continent>_<Country>_<City>/<tile_id>.json
-```
-
-The release also includes benchmark metadata:
-
-```text
-benchmark_v1/manifest_tiles_1024.csv
-benchmark_v1/manifest_patches_256.csv
-benchmark_v1/city_coverage.csv
-benchmark_v1/split_report.md
-benchmark_v1/splits/random_64_16_20/tiles_1024/{train,val,test}.txt
-benchmark_v1/splits/random_64_16_20/patches_256/{train,val,test}.txt
-```
-
-Paths in `manifest_tiles_1024.csv` describe the original unsharded layout. For the hosted WebDataset release, use `manifest_webdataset_tiles_1024.csv` for shard and member paths.
-
-## Included Cities
-
-The folder naming convention is `<Continent>_<Country>_<City>`.
-
-| Folder | Tiles |
-|---|---:|
-| `Africa_SouthAfrica_CapeTown` | 5,473 |
-| `Asia_Japan_Osaka` | 1,554 |
-| `Europe_Denmark_Aarhus` | 95 |
-| `Europe_Denmark_Copenhagen` | 1,618 |
-| `Europe_Denmark_Odense` | 90 |
-| `Europe_France_Lyon` | 101 |
-| `Europe_France_Marseille` | 159 |
-| `Europe_France_Paris` | 6,294 |
-| `Europe_France_Strasbourg` | 94 |
-| `Europe_France_Toulouse` | 231 |
-| `Europe_Germany_Berlin` | 10,355 |
-| `Europe_Germany_Frankfurt` | 99 |
-| `Europe_Germany_Munich` | 108 |
-| `Europe_Netherlands_Amsterdam` | 1,836 |
-| `NorthAmerica_Canada_Toronto` | 8,471 |
-| `NorthAmerica_Canada_Vancouver` | 126 |
-| `NorthAmerica_USA_Chicago` | 120 |
-| `NorthAmerica_USA_LosAngeles` | 123 |
-| `NorthAmerica_USA_NewYork` | 11,172 |
-| `NorthAmerica_USA_SanFrancisco` | 87 |
-| `NorthAmerica_USA_Seattle` | 78 |
-| `Oceania_Australia_Melbourne` | 602 |
-| `Oceania_Australia_Sydney` | 138 |
-| `SouthAmerica_Brazil_SaoPaulo` | 12,191 |
-
-## Splits
-
-The hosted release currently includes the `random_64_16_20` split:
-
-| Split | 1024 tiles | 256 patches | Ratio |
-|---|---:|---:|---:|
-| train | 39,178 | 626,848 | 64% |
-| validation | 9,794 | 156,704 | 16% |
-| test | 12,243 | 195,888 | 20% |
-
-These numbers sum to 61,215 hosted 1024×1024 tiles. All 256×256 patches inherit the split of their parent 1024×1024 tile to avoid leakage between train, validation, and test sets.
-
-## Data Fields
-
-- `image`: RGB image tile in JPEG format.
-- `mask`: TIFF raster mask aligned to the image tile. Pixel values encode the
-  building-height target used by the dataset generation pipeline.
-- `annotation`: COCO-style JSON annotation file for buildings in the tile.
-- `manifest_tiles_1024.csv`: one row per complete image-mask-annotation tile.
-- `manifest_patches_256.csv`: one row per 256 x 256 patch derived from the
-  1024 tile grid.
-
-## Reading and Extracting WebDataset Shards
-
-Most users do not need to manually decompress the mask TIFF files. GDAL,
-rasterio, and other TIFF readers with ZSTD-enabled libtiff support decompress
-the internal TIFF compression transparently when the array is read.
-
-Load the hosted WebDataset with Hugging Face Datasets:
-
-```python
-from datasets import load_dataset
-
-data_files = {
-    "train": "data/webdataset/train/*.tar",
-    "validation": "data/webdataset/validation/*.tar",
-    "test": "data/webdataset/test/*.tar",
-}
-
-dataset = load_dataset("webdataset", data_files=data_files, streaming=True)
-```
-
-Extract one shard with standard tar tools:
-
-```bash
-mkdir -p extracted/train
-tar -xf data/webdataset/train/train-000000.tar -C extracted/train
-```
-
-Read an extracted mask directly with rasterio:
-
-```python
-import rasterio
-
-with rasterio.open("extracted/train/Africa_SouthAfrica_CapeTown__grid_03328_z18.tiff") as src:
-    mask = src.read(1)
-```
-
-If an uncompressed TIFF file is required for a legacy tool, convert the mask
-back to an uncompressed TIFF with GDAL:
-
-```bash
-gdal_translate \
-  extracted/train/Africa_SouthAfrica_CapeTown__grid_03328_z18.tiff \
-  extracted/train/Africa_SouthAfrica_CapeTown__grid_03328_z18.uncompressed.tiff \
-  -co COMPRESS=NONE
-```
-
-To inspect the TIFF compression and checksum:
-
-```bash
-gdalinfo -checksum extracted/train/Africa_SouthAfrica_CapeTown__grid_03328_z18.tiff
-```
-
-## License and Attribution
-
-This is a multi-source geospatial dataset. The imagery and building-height
-labels are derived from public or permissioned regional sources with different
-license and attribution requirements. Do not treat the whole dataset as a
-single permissive-license corpus.
-
-See `LICENSES.md` for the source-level license inventory, attribution strings,
-and release conditions.
-
-## Intended Uses
-
-- Benchmarking monocular or single-image building height estimation.
-- Training and evaluating geospatial computer vision models.
-- Studying cross-city, cross-country, and cross-continent generalization.
-- Remote-sensing research on urban morphology and built environments.
-
-## Out-of-Scope Uses
-
-This dataset should not be used as the sole basis for legal, safety-critical,
-real-estate, insurance, tax, emergency-response, or infrastructure decisions.
-It should not be used to infer sensitive attributes about individuals or
-households.
-
-## Limitations
-
-- Coverage is uneven across continents, countries, and cities.
-- Source imagery dates and building-height label dates may not match exactly.
-- Spatial resolution, acquisition conditions, sensor characteristics, and
-  building-height definitions differ across source regions.
-- Some dense urban regions dominate the tile count.
-- The hosted `random_64_16_20` split is random at the 1024-tile level and is not a strict geographic holdout split. The full benchmark described in the paper additionally defines same-city, cross-city, and cross-country evaluation protocols.
-
-## Citation
-
-If you use HighBuild-1M, please cite the associated paper. Citation information will be updated after the paper is finalized.
-
-```bibtex
-@misc{highbuild1m2026,
-  title={HighBuild-1M: A Multi-Continental High-Resolution Benchmark Dataset for Single-View Building Height Estimation and Instance Segmentation},
-  author={Anonymous},
-  year={2026},
-  note={Submitted to NeurIPS 2026 Evaluations and Datasets Track}
-}
+data/webdataset/test/*.tar