Added reference to companion package

README.md

@@ -26,6 +26,10 @@ configs:
   data_files:
   - split: train
     path: parquet/metadata-syn/train-*.parquet
+- config_name: framewise-syn
+  data_files:
+  - split: train
+    path: parquet/framewise-syn/train-*.parquet
 - config_name: of-sta-cs
   data_files:
   - split: train
@@ -505,6 +509,136 @@ dataset_info:
   - name: train
     num_bytes: 0
     num_examples: 12000
+- config_name: framewise-syn
+  features:
+  - name: path
+    dtype: string
+  - name: dataset
+    dtype: string
+  - name: frame_labels
+    sequence:
+      dtype:
+        class_label:
+          names:
+            '0': walk
+            '1': fall
+            '2': fallen
+            '3': sit_down
+            '4': sitting
+            '5': lie_down
+            '6': lying
+            '7': stand_up
+            '8': standing
+            '9': other
+            '10': kneel_down
+            '11': kneeling
+            '12': squat_down
+            '13': squatting
+            '14': crawl
+            '15': jump
+      length: 81
+  - name: age_group
+    dtype:
+      class_label:
+        names:
+          '0': toddlers_1_4
+          '1': children_5_12
+          '2': teenagers_13_17
+          '3': young_adults_18_34
+          '4': middle_aged_35_64
+          '5': elderly_65_plus
+  - name: gender_presentation
+    dtype:
+      class_label:
+        names:
+          '0': male
+          '1': female
+  - name: monk_skin_tone
+    dtype:
+      class_label:
+        names:
+          '0': mst1
+          '1': mst2
+          '2': mst3
+          '3': mst4
+          '4': mst5
+          '5': mst6
+          '6': mst7
+          '7': mst8
+          '8': mst9
+          '9': mst10
+  - name: race_ethnicity_omb
+    dtype:
+      class_label:
+        names:
+          '0': white
+          '1': black
+          '2': asian
+          '3': hispanic_latino
+          '4': aian
+          '5': nhpi
+          '6': mena
+  - name: bmi_band
+    dtype:
+      class_label:
+        names:
+          '0': underweight
+          '1': normal
+          '2': overweight
+          '3': obese
+  - name: height_band
+    dtype:
+      class_label:
+        names:
+          '0': short
+          '1': avg
+          '2': tall
+  - name: environment_category
+    dtype:
+      class_label:
+        names:
+          '0': indoor
+          '1': outdoor
+  - name: camera_shot
+    dtype:
+      class_label:
+        names:
+          '0': static_wide
+          '1': static_medium_wide
+  - name: speed
+    dtype:
+      class_label:
+        names:
+          '0': 24fps_rt
+          '1': 25fps_rt
+          '2': 30fps_rt
+          '3': std_rt
+  - name: camera_elevation
+    dtype:
+      class_label:
+        names:
+          '0': eye
+          '1': low
+          '2': high
+          '3': top
+  - name: camera_azimuth
+    dtype:
+      class_label:
+        names:
+          '0': front
+          '1': rear
+          '2': left
+          '3': right
+  - name: camera_distance
+    dtype:
+      class_label:
+        names:
+          '0': medium
+          '1': far
+  splits:
+  - name: train
+    num_bytes: 0
+    num_examples: 12000
 - config_name: of-sta-cs
   features:
   - name: path
@@ -2115,7 +2249,7 @@ All configurations are loaded via `load_dataset("simplexsigil2/omnifall", "<conf
 - `labels` (default): All staged + OOPS labels (52k segments, 7 columns)
 - `labels-syn`: OF-Syn labels with demographic metadata (19k segments, 19 columns)
 - `metadata-syn`: OF-Syn video-level metadata (12k videos)
-- `framewise-syn`: OF-Syn frame-wise HDF5 labels (81 labels per video). **Requires the `omnifall` package (coming soon).**
+- `framewise-syn`: OF-Syn frame-wise labels (81 labels per video, from parquet)
 
 ### OF-Staged Configs
 - `of-sta-cs`: 8 staged datasets, cross-subject splits
@@ -2124,7 +2258,7 @@ All configurations are loaded via `load_dataset("simplexsigil2/omnifall", "<conf
 ### OF-ItW Config
 - `of-itw`: OOPS-Fall in-the-wild genuine accidents
 
-Video loading requires the `omnifall` package (coming soon). See examples below.
+Video loading requires the `omnifall` companion package. See examples below.
 
 ### OF-Syn Configs
 - `of-syn`: Fixed randomized 80/10/10 split
@@ -2132,7 +2266,7 @@ Video loading requires the `omnifall` package (coming soon). See examples below.
 - `of-syn-cross-ethnicity`: Cross-ethnicity split
 - `of-syn-cross-bmi`: Cross-BMI split (train: normal/underweight, test: obese)
 
-Video loading for OF-Syn configs requires the `omnifall` package (coming soon).
+Video loading for OF-Syn configs requires the `omnifall` companion package.
 
 ### Cross-Domain Evaluation
 - `of-sta-itw-cs`: Train/val on staged CS, test on OOPS
@@ -2192,15 +2326,46 @@ syn_labels = load_dataset("simplexsigil2/omnifall", "labels-syn")["train"]
 
 ### Loading Videos
 
-Video loading (OF-Syn, OF-ItW, and cross-domain configs) requires the `omnifall` Python package, which will be available on PyPI soon. The package handles video download, caching, and integration with HuggingFace datasets.
+Video loading requires the `omnifall` companion package (`pip install omnifall`), which handles video download, caching, and path resolution.
+
+```python
+import omnifall
+
+# OF-Syn: videos are auto-downloaded from HF Hub (~9.1GB, cached)
+ds = omnifall.load("of-syn", video=True)
+
+# OF-ItW: requires one-time OOPS video preparation (~45GB stream, ~2.6GB on disk)
+omnifall.prepare_oops()  # interactive license consent
+ds = omnifall.load("of-itw", video=True)
+
+# Cross-domain: syn videos for train/val, OOPS for test
+ds = omnifall.load("of-syn-itw", video=True)
+
+# The video column contains absolute file paths (strings)
+print(ds["train"][0]["video"])  # /home/user/.cache/omnifall/of-syn-videos/fall/fall_to_001.mp4
+```
 
-For OOPS videos specifically, you can prepare them manually using the included script:
+The `video` column contains string paths rather than decoded video frames, since temporal segmentation tasks require frame-level control via PyAV or decord.
+
+You can also add video paths to an already-loaded dataset:
+
+```python
+from datasets import load_dataset
+ds = load_dataset("simplexsigil2/omnifall", "of-syn")
+ds = omnifall.add_video(ds, config="of-syn")
+```
+
+OOPS videos can also be prepared via CLI or the standalone script:
 
 ```bash
+# Via omnifall CLI
+omnifall prepare-oops
+
+# Via standalone script (alternative)
 python prepare_oops_videos.py --output_dir /path/to/oops_prepared
 ```
 
-The preparation streams the full OOPS archive from the original source and extracts only the 818 videos used in OF-ItW. The archive is streamed and never written to disk, so only ~2.6GB of disk space is needed. If you already have the OOPS archive downloaded locally, pass it with `--oops_archive /path/to/video_and_anns.tar.gz`.
+The preparation streams the full OOPS archive from the original source and extracts only the 818 videos used in OF-ItW. The archive is streamed and never written to disk, so only ~2.6GB of disk space is needed. If you already have the OOPS archive downloaded locally, pass it with `--oops-archive /path/to/video_and_anns.tar.gz`.
 
 ## Label definitions
 

generate_parquet.py

@@ -9,6 +9,9 @@ Usage:
 """
 
 import os
+import subprocess
+import tarfile
+import tempfile
 from pathlib import Path
 
 import numpy as np
@@ -322,6 +325,75 @@ def gen_individual(ds_name):
     return {ds_name: results}
 
 
+def gen_framewise_syn():
+    """Config: framewise-syn - OF-Syn frame-wise labels from HDF5, single train split.
+
+    Reads syn_frame_wise_labels.tar.zst (HDF5 files with 81 frame labels each),
+    merges with videos/metadata.csv for demographics, writes a single parquet.
+    Requires h5py at generation time only.
+    """
+    import h5py
+
+    archive_path = REPO_ROOT / "data_files" / "syn_frame_wise_labels.tar.zst"
+    if not archive_path.exists():
+        print(f"  SKIP framewise-syn: archive not found at {archive_path}")
+        return {"framewise-syn": {}}
+
+    metadata_df = load_csv(METADATA_FILE)
+
+    metadata_fields = DEMOGRAPHIC_COLUMNS
+
+    rows = []
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        # Extract the .tar.zst archive (tarfile doesn't support zstd natively)
+        subprocess.run(
+            ["tar", "--zstd", "-xf", str(archive_path), "-C", tmpdir],
+            check=True,
+        )
+        tmppath = Path(tmpdir)
+        h5_files = sorted(tmppath.glob("**/*.h5"))
+        print(f"  Found {len(h5_files)} HDF5 files in archive")
+
+        for h5_file_path in h5_files:
+            relative_path = h5_file_path.relative_to(tmppath)
+            video_path = str(relative_path.with_suffix(""))
+
+            try:
+                with h5py.File(h5_file_path, "r") as f:
+                    frame_labels = f["label_indices"][:].tolist()
+            except Exception as e:
+                print(f"  WARNING: Failed to read {h5_file_path}: {e}")
+                continue
+
+            video_metadata = metadata_df[metadata_df["path"] == video_path]
+            if len(video_metadata) == 0:
+                print(f"  WARNING: No metadata for {video_path}, skipping")
+                continue
+            video_meta = video_metadata.iloc[0]
+
+            row = {
+                "path": video_path,
+                "dataset": "of-syn",
+                "frame_labels": frame_labels,
+            }
+            for field in metadata_fields:
+                if field in video_meta and pd.notna(video_meta[field]):
+                    row[field] = str(video_meta[field])
+                else:
+                    row[field] = ""
+            rows.append(row)
+
+    df = pd.DataFrame(rows)
+    # Cast demographic columns to string
+    df = cast_demographic_dtypes(df)
+    df["path"] = df["path"].astype(str)
+    df["dataset"] = df["dataset"].astype(str)
+
+    path = write_parquet(df, "framewise-syn", "train")
+    return {"framewise-syn": {"train": len(df)}}
+
+
 # ---- Main ----
 
 def main():
@@ -332,7 +404,7 @@ def main():
 
     # Labels configs (single train split)
     print("\n--- Labels configs ---")
-    for gen_fn in [gen_labels, gen_labels_syn, gen_metadata_syn]:
+    for gen_fn in [gen_labels, gen_labels_syn, gen_metadata_syn, gen_framewise_syn]:
         result = gen_fn()
         all_results.update(result)
         for config, splits in result.items():

omnifall_dataset_examples.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:80442a8e82cee61a8c64d78b081b9c9271438cc7442a1ab3fcd12dcd472537f3
-size 10103509
+oid sha256:5b52439f11c721e01c5db072454e4c7571bdd2a72df9f04acbf62b6735f218c0
+size 10085652

parquet/framewise-syn/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2c88f22010e64afa58b3ee0d7ddacf84525386aed594346ecc0680504f976cd
+size 139176