scales/scale_comparison.md

# EventActivityNet Scale Comparison Audit

Status: final scaling verification audit for the candidate Large / Medium / Small release manifests. No manifests or H5 files were modified.

## Executive Conclusion

- The recovered Large manifest exactly matches `the recovered original Large subset manifest` by video ID.
- `Small` is a strict subset of `Medium`, and `Medium` is a strict subset of `Large`.
- All three scales preserve all 200 action classes present in the recovered Large release.
- The scales are suitable for public release as candidate manifests, with the caveats listed below.

## Part A - Large Provenance Verification

### Confirmed By Code

| Principle | Evidence |
|---|---|
| Train/validation merge before sampling | `load_metadata(paths)` merges every provided metadata file into one `meta` dict before class sampling. The sampling loop iterates over `metadata.keys()` without split-specific quotas. |
| Seed | `random.Random(args.seed)` and `np.random.seed(args.seed)`; default `--seed 2025`. |
| Minimum-per-class rule | Initial class-balanced stage uses `k = max(5, int(args.class_ratio * len(vids)))`; default `class_ratio=0.2`. |
| Duration stratification | `length_bucket_edges(durs)` uses `np.quantile(durations, [0.33, 0.66])`. |
| Event-friendly definition | `event_friendly()` returns caption keyword match OR first-frame mean below threshold. |
| Event-friendly keywords | `run, fast, sprint, night, dark, slow-motion` |
| Darkness threshold | default `--brightness-threshold 0.4`. |

### Confirmed By Data

- Candidate train/validation pool: 14,926 videos, 200 classes.
- Reproduced class-balanced stage: 2,908 videos, 200 classes.
- Reproduced length-balanced stage: 2,888 videos, 200 classes.
- Original sample_subset caption-duration quantiles: q33=76.306200s, q66=151.323400s.
- Large equals `subset_caps_20p.json` by ID: `True`.
- Large split counts: train 2,316, validation 947.
- Large event-friendly ratio from manifest fields: 0.653080.
- Event flags are consistent with `keyword OR dark_first_frame` fields: `True`.
- Final Large classes below the initial seed-stage per-class quota: 54. This is a consequence of later global length balancing, not evidence of a different subset.

### Inferred

- The released Large subset satisfies the recovered curation principles as the final output of `sample_subset.py`: it matches the original subset list exactly, preserves all 200 local classes, and reflects the expected duration/event/class-balanced structure.
- Medium and Small preserve the Large release principles by deterministic nested stratification over `(split, class_label, duration_bucket, event_friendly)`, but they are newly generated derivatives rather than historical `sample_subset.py` outputs.

### Deviations / Historical Inconsistencies

- **final_large_not_constrained_by_initial_min_per_class_rule**: `sample_subset.py` enforces `max(5, int(class_ratio * class_count))` only during the initial class-balanced stage. The later length-balancing stage samples globally from that pool, so the final Large manifest is not expected to retain those initial per-class quotas. Affected classes: 54; largest shortfalls: Swimming 6/15, Shaving 6/14, Baton twirling 13/20, Cleaning windows 8/14, Shoveling snow 9/15, Table soccer 8/14, Archery 9/14, Clipping cat claws 8/13.
- **release_scale_duration_buckets_use_src_fmt_dur_not_original_caption_metadata_duration**: The scale manifests use src_fmt_dur from verify_all.csv per the scale-construction requirement. sample_subset.py used metadata[v]["duration"] for its original length balancing.
- **class_count_claim_203_not_supported_by_local_sources**: Local taxonomy/candidate/final files contain and preserve 200 classes; see audit/missing_classes_analysis.md.

## Part B - Scale Comparison

| Scale | Videos | Hours | Avg dur (s) | Median dur (s) | Train | Validation | Classes | Event-friendly |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| Large | 3,263 | 106.941600381 | 117.986 | 114.242 | 2,316 | 947 | 200 | 0.653080 |
| Medium | 1,537 | 50.000000128 | 117.111 | 114.056 | 1,074 | 463 | 200 | 0.648016 |
| Small | 667 | 20.000000374 | 107.946 | 102.864 | 473 | 194 | 200 | 0.646177 |

### Duration Bucket Distribution

| Scale | Short | Medium | Long | Short ratio | Medium ratio | Long ratio |
|---|---:|---:|---:|---:|---:|---:|
| Large | 1,077 | 1,076 | 1,110 | 0.330064 | 0.329758 | 0.340178 |
| Medium | 512 | 500 | 525 | 0.333116 | 0.325309 | 0.341574 |
| Small | 247 | 226 | 194 | 0.370315 | 0.338831 | 0.290855 |

### Per-Class Deviation From Large

| Scale | Max ratio deviation | Mean ratio deviation | Missing Large classes |
|---|---:|---:|---:|
| Large | 0.000000000 | 0.000000000 | 0 |
| Medium | 0.003145020 | 0.000548689 | 0 |
| Small | 0.006468877 | 0.001455918 | 0 |

### Largest Class Gains / Losses

#### Medium vs Large

Largest gains by class ratio:
- Snowboarding: delta 0.001640, medium count 11, Large count 18
- Building sandcastles: delta 0.001603, medium count 10, Large count 16
- Clean and jerk: delta 0.001565, medium count 9, Large count 14
- Fixing bicycle: delta 0.001527, medium count 8, Large count 12
- Cleaning sink: delta 0.001452, medium count 6, Large count 8

Largest losses by class ratio:
- Removing ice from car: delta -0.003145, medium count 6, Large count 23
- Making a cake: delta -0.001957, medium count 5, Large count 17
- Mooping floor: delta -0.001919, medium count 6, Large count 19
- Cutting the grass: delta -0.001613, medium count 6, Large count 18
- Sumo: delta -0.001575, medium count 7, Large count 20

#### Small vs Large

Largest gains by class ratio:
- Washing dishes: delta 0.004059, small count 7, Large count 21
- Discus throw: delta 0.004026, small count 8, Large count 26
- Sharpening knives: delta 0.003819, small count 5, Large count 12
- Wrapping presents: delta 0.003786, small count 6, Large count 17
- Table soccer: delta 0.003545, small count 4, Large count 8

Largest losses by class ratio:
- Skiing: delta -0.006469, small count 1, Large count 26
- Capoeira: delta -0.004663, small count 2, Large count 25
- Pole vault: delta -0.003777, small count 3, Large count 27
- Cheerleading: delta -0.003744, small count 2, Large count 22
- Playing harmonica: delta -0.003711, small count 1, Large count 17

### Nestedness

- Small strict subset of Medium: `True`
- Medium strict subset of Large: `True`
- All IDs unique per scale: `{'large': True, 'medium': True, 'small': True}`

### Published Paper Duration Comparison

Published Large duration: 107.3 h.

| Duration source | Hours | Difference from 107.3 h |
|---|---:|---:|
| Release manifest `src_fmt_dur` | 106.941600381 | -0.358399619 |
| `subset_caps_20p.json` caption duration | 106.952283333 | -0.347716667 |
| `verify_all.csv` `src_dur` | 106.848348413 | -0.451651587 |
| `verify_all.csv` `src_fmt_dur` | 106.941600381 | -0.358399619 |
| `verify_all.csv` `src_dur_est` | 106.782090037 | -0.517909963 |
| `verify_all.csv` `dst_dur` | 106.785122166 | -0.514877834 |
| `verify_all.csv` `dst_fmt_dur` | 106.939227018 | -0.360772982 |
| `verify_all.csv` `dst_dur_est` | 106.781256732 | -0.518743268 |

The recovered Large IDs exactly match subset_caps_20p.json, so the observed 107.3 h discrepancy does not indicate a different historical subset. It is not explained solely by src_fmt_dur versus other local metadata fields; all local fields are around 106.78-106.95 h. Treat 107.3 h as an approximate/reporting duration unless another historical duration source is recovered.

## Part C - Release Recommendation

Recommendation: **Suitable as candidate public release manifests after documenting caveats.**

Remaining caveats:
- Use 200 classes, not 203, for this recovered release.
- Medium and Small are newly generated deterministic nested candidates; they are distribution-preserving derivatives of Large, not historical sample_subset.py outputs.
- Authoritative scale durations use src_fmt_dur from verify_all.csv. The original subset file caption durations and all local verify_all duration fields are close but do not equal the paper statement of 107.3 h exactly.
- The final Large list can have classes below the initial max(5, int(class_ratio * class_count)) seed-stage minimum because the original script did not enforce that constraint after length balancing.

Machine-readable details are in `scale_comparison.json`.