docs/DATASET_FORMAT.md

# EventActivityNet Dataset Format

## Representations

EventActivityNet stores one HDF5 file per video for each supported bin count.

| Bin count `B` | Public payload | Event shape |
|---:|---|---|
| 5 | `data_5bin/` | `(T5, 5, H, W)` |
| 9 | `data_9bin/` | `(T9, 9, H, W)` |

File size varies substantially with duration and spatial resolution.

## HDF5 Schema

Every HDF5 file has exactly these root datasets:

```text
events
voxel_event_start
voxel_event_count
```

### `events`

| Property | Value |
|---|---|
| Shape | `(T_B, B, H, W)` |
| Dtype | `int16` |
| Compression | gzip, level 4 |
| Shuffle | enabled |
| Chunking | `(1, B, min(H, 256), min(W, 256))` |

### `voxel_event_start`

| Property | Value |
|---|---|
| Shape | `(T_B,)` |
| Dtype | `int64` |
| Compression | LZF |
| Shuffle | enabled |
| Chunking | `(1024,)` |

### `voxel_event_count`

| Property | Value |
|---|---|
| Shape | `(T_B,)` |
| Dtype | `int32` |
| Compression | LZF |
| Shuffle | enabled |
| Chunking | `(1024,)` |

Required root attributes are `fps`, `height`, `width`, `num_bins`, and
`interpolate_bins`. `num_bins` is 5 or 9 and matches `events.shape[1]`.

## Transition Grouping

Construction follows decoded frame order. For `N` source frames there are
`N - 1` adjacent-frame transitions. Transition index `e` corresponds to source
frames `(e, e + 1)`.

For bin count `B`:

```text
T_B = ceil((N - 1) / B)
voxel_event_start[t] = B * t
voxel_event_count[t] = min(B, N - 1 - B * t)
```

`events[t]` groups transitions in the half-open range
`[B*t, min(B*t + B, N - 1))`. The associated source-frame interval is
`[B*t, min(B*t + B, N - 1)]`. The final group can contain fewer than `B`
valid transitions; unused bins are zero-filled.

## Timing

The HDF5 `fps` attribute is source-frame FPS stored as a float. Use the released
`fps_num` and `fps_den` fields for reproducible conversion. Approximate voxel
times are:

```text
start_seconds = B * t * fps_den / fps_num
end_seconds = min(B * t + B, N - 1) * fps_den / fps_num
```

Per-frame presentation timestamps are not consumed. These conversions are
therefore approximate for within-video variable-frame-rate streams. Do not
assume fixed 25 fps or 240 fps, and do not use `t / fps` as voxel time.

For a caption/action interval `[start_seconds, end_seconds]`:

```text
start_frame = floor(start_seconds * fps_num / fps_den)
end_frame = ceil(end_seconds * fps_num / fps_den)
t_start = max(0, floor(start_frame / B))
t_end_exclusive = min(T_B, ceil(end_frame / B))
```

Use `[t_start, t_end_exclusive)` for Python slicing.

## Memory-Safe Loading

```python
import h5py

with h5py.File("v_example.h5", "r") as f:
    events = f["events"]
    starts = f["voxel_event_start"]
    counts = f["voxel_event_count"]

    B = int(f.attrs["num_bins"])
    print(events.shape)  # (T_B, B, H, W)
    print(events.dtype)  # int16
    print(starts.dtype)  # int64
    print(counts.dtype)  # int32

    selected = events[10:18]  # reads only the selected temporal range
```

Avoid loading complete event tensors unless sufficient memory is available.

## Shards and Metadata

Each representation has 157 train shards, 62 validation shards, and 3,263 HDF5
members. Representation-specific manifests and checksums are under:

```text
metadata/5bin/
metadata/9bin/
```

Shared source/timing metadata is in `metadata/video_metadata.jsonl`, with
representation-specific tensor fields nested under `representations`.