# EventActivityNet Annotation and Original-Rate Timing Metadata

This package contains public annotation and timing metadata for the EventActivityNet v1.0 Large release. It is filtered to the canonical 3,263 release video IDs. Medium and Small scales reuse these same files by filtering with their existing video-ID lists.

## Files

- `annotations/activitynet_captions.json`: ActivityNet Captions entries for release videos. Training videos have the `train` caption set; validation videos may include separate `val_1` and `val_2` caption sets.
- `annotations/activitynet_actions.json`: ActivityNet v1.3 action annotations for release videos.
- `annotations/eventactivitynet_alignment.json`: Project-generated action-caption alignment. The alignment was generated with temporal IoU between action segments and caption timestamps, with midpoint-distance fallback, using `align_v13_with_captions.py`.
- `annotations/annotation_issues.jsonl`: Known source annotation quirks preserved without modification.
- `metadata/video_metadata.jsonl`: Original-rate production timing metadata for each release video.

## Timing

EventActivityNet v1.0 uses original-rate, variable-FPS source videos. It was not generated from the interpolated 240 fps video derivatives. Caption and action timestamps remain in seconds on the original ActivityNet video timeline.

For a video with `source_frame_count = N`, `num_bins = 5`, and `frames_per_bin = 1`, the production tensor length is:

```text
events_T = ceil((N - 1) / 5)
```

Voxel index `i` accumulates adjacent-frame event slices beginning at slice `5*i`. Its approximate source-frame interval is `[5*i, min(5*i + 5, N - 1)]`, and its approximate seconds interval is `[5*i / fps, min(5*i + 5, N - 1) / fps]`. Use the exact rational FPS fields where available.

## Known Source Annotation Quirks

Canonical source annotations are preserved as-is. Some source timestamps or action segments have minor ordering or boundary issues, including small floating-point overshoots beyond the recorded duration. Consumers should use a small numerical tolerance when validating temporal boundaries; the issue file records the exact observed cases and overshoot values.

## Source Hashes

Source SHA256 hashes used to build this package:

```json
{
  "activitynet_actions": "4c29d5b1561e1cbff9ac69816c159e60d417a18f16db8acc0fc254c377fa9ae3",
  "alignment": "40a75cfd6a053011766e80e5a6d449b84a3ececf51b1f34cfa33eefba8729d36",
  "captions_train": "704536b2779011c1a729c927ae30e39ea9c90c38557836bbe969499917243d69",
  "captions_val_1": "466aaccb9ebe45c6db4c98488a4ddbeb5b22faffe62e1bbce81222c14fc29e1c",
  "captions_val_2": "08fe56568bc92dd02a6150fe1094457a5f2ac72975b8778a922bc9a65aa7f0e9",
  "class_mapping": "f15c7864abe012bd82b1df2c33beae8254acffc4b86a3dcbf6174c02a61eb400",
  "large_manifest": "02468b7f301a85c5f3776b72b9b1ff83635bc99dd6b2ce42a7c7e5308007c96a",
  "verify_all": "a5d17bcb508789c6e96747d3eac9ecee13b99dd683df995cd81e3179966bd7bc"
}
```