# LegibleTCN results compared with published action-recognition systems

Research checked 15 September 2026. Measurements: the [completed run report](/action-benchmark-run.md), [machine-readable results](/action-benchmark-results.json), and [checksum-verified executed source with line references](/benchmarks/2026-09-15/architecture-source.html).

## Conclusion

**The current classifier contains 25,680–32,664 parameters, with recognition accuracy substantially below strong published systems.** Its purpose is inexpensive local recognition: use a small action classifier on body poses, show predictions, and let people use or correct them to grow a training set. CPU execution in browsers and on phones is the deployment target; this run measured accuracy and model size, without measuring device latency.

The most useful comparison is with models using the same released HRNet 2D poses. Comparing only against large RGB models would obscure a substantial gap that already exists within the pose modality. Recent skeleton ensembles and large video models provide additional context, but their inputs, training and inference budgets differ.

“Published reference” below means a result reported by the method's authors or its official implementation. These references were not rerun locally, and the research does not establish an exhaustive, independently reproduced September 2026 leaderboard.

## 1. Your measured results

All results below are **pre-cut clip classification**, using one actor, 64 resampled frames and a noncausal classifier. Every protocol trained three seeds from scratch for 80 epochs; final-epoch models were evaluated without test-set checkpoint selection. Standard deviations describe variation between seeds, not statistical confidence intervals or run-to-run data uncertainty.

| Dataset and protocol | Your top-1 accuracy | Your mean-class accuracy | Parameters |
|---|---:|---:|---:|
| HMDB51, three official splits | 34.19 ± 0.38% | 34.19% | 25,971 |
| UCF101, three official splits | 50.60 ± 0.09% | 50.29% | 30,821 |
| Diving48, released PYSKL train/test | 9.05 ± 0.66% | 6.75% | 25,680 |
| FineGym Gym99 | 74.11 ± 0.14% | **68.55%** | 30,627 |
| NTU60, cross-subject, HRNet 2D | 76.39 ± 0.19% | 76.38% | 26,844 |
| NTU60, cross-view, HRNet 2D | 82.28 ± 0.10% | 82.28% | 26,844 |
| NTU120, cross-subject, HRNet 2D | 64.58 ± 0.38% | 66.44% | 32,664 |
| NTU120, cross-setup, HRNet 2D | 66.04 ± 0.51% | 66.00% | 32,664 |

HMDB51/UCF101 average each seed over the three official splits before averaging seeds. Gym99 comparisons must use **68.55% mean-class accuracy**, which gives each class equal weight; its 74.11% sample-weighted accuracy is a different metric. Diving48's corrected V2 label correspondence remains unverified, and its 1,970-example test subset contains 47 represented classes.

## 2. Closest established comparison: released HRNet 2D poses

The following PoseC3D joint-stream scores come from the official PYSKL model zoo. They share the pose source and named protocols with this run, but use a larger model and multiple test clips/flip augmentation. The cited NTU60 recipe trains for 24 epochs of a 10× repeated dataset, equivalent to 240 dataset passes. They therefore measure a practical accuracy gap, not an isolated architecture experiment. [PYSKL PoseC3D model zoo](https://github.com/kennymckormick/pyskl/blob/f2bf3a6b08e2e8dec744692d64efdb187fd6719a/configs/posec3d/README.md), [exact NTU60 training configuration](https://github.com/kennymckormick/pyskl/blob/f2bf3a6b08e2e8dec744692d64efdb187fd6719a/configs/posec3d/slowonly_r50_ntu60_xsub/joint.py#L59-L74)

| Protocol and metric | Your result | PoseC3D joint stream | Your shortfall, percentage points |
|---|---:|---:|---:|
| NTU60 cross-subject, top-1 | 76.39% | 93.7% | 17.31 |
| NTU60 cross-view, top-1 | 82.28% | 96.5% | 14.22 |
| NTU120 cross-subject, top-1 | 64.58% | 85.9% | 21.32 |
| NTU120 cross-setup, top-1 | 66.04% | 89.7% | 23.66 |
| Gym99, mean-class | 68.55% | 93.8% | 25.25 |

The same official release reports Kinetics-pretrained PoseC3D at 69.4% on **HMDB51 split 1** and 86.9% on **UCF101 split 1**. Your matching split-1 results are 33.07% and 49.97%, respectively: shortfalls of 36.33 and 36.93 points. Your three-split aggregate must not be presented as the matching split-1 result. [PYSKL split and pretraining notes](https://github.com/kennymckormick/pyskl/blob/main/configs/posec3d/README.md)

An additional coordinate-based reference is ST-GCN++. Its official **HRNet 2D joint-only** scores are 89.3/97.4% on NTU60 cross-subject/cross-view and 84.4/88.1% on NTU120 cross-subject/cross-setup. Four-stream fusion reaches 93.2/98.5% and 86.4/90.3%. This confirms that the gap is not solely attributable to converting coordinates into heatmaps. [Official ST-GCN++ results](https://github.com/kennymckormick/pyskl/blob/main/configs/stgcn%2B%2B/README.md)

## 3. Recent high-performing research

### Skeleton recognition

| Method | Publication status | NTU60 X-Sub / X-View | NTU120 X-Sub / X-Set | Gym99 mean-class |
|---|---|---:|---:|---:|
| ProtoGCN, six-model ensemble | CVPR 2025 | 93.8 / 97.8% | 90.9 / 92.2% | 95.9% |
| SkeletonAgent, six-model ensemble | November 2025 preprint | 94.5 / 98.2% | 91.7 / 93.1% | 96.5% |

ProtoGCN's NTU experiments use the native 25-joint skeleton setting, while its Gym99 experiment uses PYSKL poses. Its Gym99 score is 27.35 points above yours. [ProtoGCN paper, Tables 1–2 and dataset protocols](https://arxiv.org/html/2411.18941v2)

SkeletonAgent also uses the native NTU skeleton setting; its Gym99 data comes from PYSKL. Its training incorporates language guidance from CLIP and an LLM, and its highest reported scores combine six models. Its Gym99 score is 27.95 points above yours. These are contemporary high-performing references, not matched 2D-NTU single-classifier comparisons. [SkeletonAgent paper, Table 1 and Sections 4.1–4.3](https://arxiv.org/html/2511.22433v1)

### RGB and multimodal video recognition

| Method | HMDB51 | UCF101 | Important distinction |
|---|---:|---:|---|
| Your LegibleTCN | 34.19% | 50.60% | HRNet poses; classifier trained from scratch |
| VideoMAE V2, ViT-giant | 88.1% | 99.6% | RGB; large-scale pretraining, K710 intermediate training, 5 clips × 3 crops |
| FTP-UniFormerV2-L/14 | 91.5% | 99.7% | RGB; VLM-assisted training; reported 402M parameters |
| OmniVec2 | 92.1% | 99.7% | Broad multimodal pretraining; reported three-split evaluation |

VideoMAE V2's authors provide the scores and inference setup in their [official model zoo](https://github.com/OpenGVLab/VideoMAEv2/blob/master/docs/MODEL_ZOO.md). The more recent FTP journal article has a [2026 publication record](https://www.sciencedirect.com/science/article/abs/pii/S0031320326007594); its [author manuscript](https://arxiv.org/html/2403.16128v1) explicitly describes three-split averaging and the figures above. OmniVec2's authors report the higher HMDB51 figure in [their CVPR 2024 supplement, Tables 5–6](https://openaccess.thecvf.com/content/CVPR2024/supplemental/Srivastava_OmniVec2_-_A_CVPR_2024_supplemental.pdf).

The broad video references are roughly 49 points ahead on UCF101 and 54–58 points ahead on HMDB51. Their access to appearance, objects and larger pretrained representations makes them a different engineering choice. Those differences do not establish how much of the accuracy gap is caused by each design decision.

### Diving48 needs a protocol audit before a ranked comparison

Your 9.05% is a weak result within the downloaded split. The PYSKL PoseC3D release lists 54.5%, but label-version equivalence remains unverified. It should not be used to manufacture a precise “gap to V2 SOTA.” [Official PYSKL result](https://github.com/kennymckormick/pyskl/blob/main/configs/posec3d/README.md)

For context, the August 2026 **FineX preprint** reports 92.9% on Diving48 using RGB and pose features. It describes approximately 16.1K training and 2.3K test clips and uses D-FINE plus ViTPose-L for Diving48. Our downloaded PYSKL release has 15,027/1,970 examples and HRNet poses. This is both a data-coverage and an input-pipeline mismatch. FineX also reports 96.4% Gym99 mean-class accuracy and 97.1% sample accuracy, reinforcing why these two metrics must remain distinct. [FineX paper, Section 4.1 and Table 1](https://arxiv.org/html/2608.13458v1)

## 4. Small models for local recognition

The fresh classifiers contain **25.7K–32.7K parameters**, with approximately **100–128 KiB of raw FP32 weights** and **105–133 KiB serialized checkpoints**. These are measured counts and file sizes for individual seeds. They exclude the pose estimator, human detector, tracking, runtime libraries and buffers.

The official MMAction2 implementation lists approximately 2.0M parameters for its NTU60/FineGym PoseC3D configurations. Your corresponding 26,844/30,627-parameter classifiers have roughly **75×/65× fewer parameters**. This comparison describes action classifiers, and the listed official checkpoints are not exactly the PYSKL checkpoints used for the score table above. It is not an end-to-end system-size or speed comparison. [MMAction2 PoseC3D configurations and sizes](https://github.com/open-mmlab/mmaction2/blob/main/configs/skeleton/posec3d/README.md)

The intended workflow is to stream video into a browser or phone, extract body poses, and run the compact action classifier locally on the CPU. People can use the predictions immediately or review and correct their labels and boundaries, adding examples to a growing dataset. That makes the small model useful as a first stage toward training heavier methods with more varied, reliable supervision. Corrected misses, false alarms and background examples matter alongside successful detections; saving unreviewed predictions would repeat the model's existing errors.

## 5. What the results justify doing next

1. **Run the continuous benchmarks under strictly causal processing.** Pre-cut scores cannot answer when an action starts, whether the system produces false alarms during background activity, or how late a decision arrives. Archived causal-network scores also do not establish a causal preprocessing pipeline.
2. **Verify Diving48's corrected labels and exclusions.** Until then, retain the measured result with its caveat and exclude it from a ranked V2 comparison.
3. **Use a matched baseline to locate the accuracy loss.** Same split, same poses, same person selection, same training budget and same test sampling; compare a compact reference classifier with LegibleTCN. The current measurements alone cannot identify the cause of the loss.
4. **Measure the actual product tradeoff.** Report end-to-end latency, pose-estimator cost, memory and accuracy on the same hardware. Small weights alone do not prove a faster complete system.

Possible contributors to investigate include discarding other actors, coordinate normalization, fixed-length resampling and limited temporal capacity. These are hypotheses from the current implementation, not established explanations of the observed gaps.
