# LegibleTCN action recognition run — 2026-09-15

Completed 12 split protocols across six dataset families, with three seeds each (36 trained classifiers). These are pre-cut clip classification results, not continuous detection or streaming scores.

## Results

| Dataset | Protocol | Top-1 % ± seed SD | Mean-class % | Parameters | Checkpoint bytes |
|---|---|---:|---:|---:|---:|
| HMDB51 | 3 official splits | 34.19 ± 0.38 | 34.19 | 25971 | 109105 |
| UCF101 | 3 official splits | 50.60 ± 0.09 | 50.29 | 30821 | 128497 |
| Diving48 | PYSKL train / test · V2 labels not yet cross-checked | 9.05 ± 0.66 | 6.75 | 25680 | 107993 |
| FineGym Gym99 | Train / validation | 74.11 ± 0.14 | 68.55 | 30627 | 127669 |
| NTU RGB+D 60 | Cross-subject · 2D HRNet | 76.39 ± 0.19 | 76.38 | 26844 | 112541 |
| NTU RGB+D 60 | Cross-view · 2D HRNet | 82.28 ± 0.10 | 82.28 | 26844 | 112661 |
| NTU RGB+D 120 | Cross-subject · 2D HRNet | 64.58 ± 0.38 | 66.44 | 32664 | 135857 |
| NTU RGB+D 120 | Cross-setup · 2D HRNet | 66.04 ± 0.51 | 66.00 | 32664 | 135957 |

HMDB51 and UCF101 average the three official splits separately for each seed, then report the mean and population standard deviation of those three seed averages. Per-split values are below. File size is for one seed, not an ensemble.

## Fixed recipe

Released PYSKL HRNet COCO-17 poses; first actor in original order; image-centered coordinates and confidence masking at 0.01; x/y/confidence/dx/dy channels; 64-frame uniform resampling; 8 filters per joint, 96 hidden channels; noncausal LegibleTCN; 80 epochs; batch 64; AdamW with learning rate 0.002, weight decay 0.0001, cosine schedule; class-balanced replacement sampling and label smoothing 0.1; no augmentation. Seeds 42, 43, 44. Final epoch, with no test-set checkpoint selection. Fresh initialization for every split and seed.

Classifier training ran on an RTX 3090. This does not include pose-extraction cost, parameters or runtime. Released HRNet poses are not an evaluation of the application's current pose extractor. Parameter bytes are exact counts × 4; checkpoint bytes include PyTorch serialization. No latency or human-legibility score was measured.

## Split receipts

| Protocol and complete recipe | Train clips | Test clips | Top-1 % ± seed SD |
|---|---:|---:|---:|
| [HMDB51 split 1](/benchmarks/2026-09-15/hmdb51.json) | 3570 | 1530 | 33.07 ± 1.05 |
| [HMDB51 split 2](/benchmarks/2026-09-15/hmdb51_split2.json) | 3570 | 1530 | 35.64 ± 0.58 |
| [HMDB51 split 3](/benchmarks/2026-09-15/hmdb51_split3.json) | 3570 | 1530 | 33.86 ± 0.30 |
| [UCF101 split 1](/benchmarks/2026-09-15/ucf101.json) | 9537 | 3783 | 49.97 ± 0.62 |
| [UCF101 split 2](/benchmarks/2026-09-15/ucf101_split2.json) | 9586 | 3734 | 51.29 ± 0.57 |
| [UCF101 split 3](/benchmarks/2026-09-15/ucf101_split3.json) | 9624 | 3696 | 50.52 ± 0.33 |
| [Diving48](/benchmarks/2026-09-15/diving48.json) | 15027 | 1970 | 9.05 ± 0.66 |
| [FineGym Gym99](/benchmarks/2026-09-15/gym.json) | 20484 | 8521 | 74.11 ± 0.14 |
| [NTU RGB+D 60 X-Sub (2D HRNet)](/benchmarks/2026-09-15/ntu60.json) | 40091 | 16487 | 76.39 ± 0.19 |
| [NTU RGB+D 60 X-View (2D HRNet)](/benchmarks/2026-09-15/ntu60_xview.json) | 37646 | 18932 | 82.28 ± 0.10 |
| [NTU RGB+D 120 X-Sub](/benchmarks/2026-09-15/ntu120.json) | 63026 | 50919 | 64.58 ± 0.38 |
| [NTU RGB+D 120 X-Set (2D HRNet)](/benchmarks/2026-09-15/ntu120_xset.json) | 54468 | 59477 | 66.04 ± 0.51 |

The adapter requires disjoint training/test identities and complete coverage of the released split lists. Dataset and source hashes, execution logs and weights are retained locally in artifacts/action-benchmarks. Original benchmark media exclusions and pose-estimation pretraining are not independently audited. Diving48's corrected V2 label correspondence remains explicitly unverified.

## Execution record

The initial job completed HMDB51 and UCF101, then stopped on Diving48 because the released key is test, not val. The adapter was corrected after inspecting the pickle schema, then execution continued from Diving48 without rerunning or tuning the completed families. This metadata fix does not change the model or training recipe.

## Scope still outstanding

Kinetics needs separate pose shards. Native 3D datasets, BABEL-TAL, OAD and other continuous benchmarks require data access and adapters/evaluators. The archived PKU-MMD scores are historical and were not rerun by this clip suite. See the [benchmark landscape](/action-recognition-benchmark-landscape.md) and [historical measurements](/benchmark-results.md).

## Class coverage

Diving48 also exposed absent classes in its test split. Mean-class accuracy averages recall over classes with test examples; it does not divide by zero or invent scores for unobserved classes. Per-class support and correct counts are saved in the later split receipts. The earlier HMDB51/UCF101 runs contain every class and their metric is unchanged. This correction only affects the reporting step, not training.
