# FSJump6: measured few-shot recognition

Completed 15 September 2026. These are measured runs on supplied action cuts, not published scores or live segmentation results. Detection and retrieval comparisons are still in progress.

## Results

Top-1 accuracy, mean ± population standard deviation across five supplied support sets. Every trial evaluates the same 520 held-out actions. Four/eight examples per class means 24/48 labeled actions total.

| Method | Four examples per class | Eight examples per class | Raw results |
| --- | ---: | ---: | --- |
| LegibleTCN, 80 epochs | 17.15 ± 6.94% | 19.69 ± 4.37% | [JSON](fs-recognition-legible-tcn.json) |
| LegibleTCN, separate 2,000-epoch run | 20.92 ± 4.24% | 26.92 ± 2.59% | [JSON](fs-recognition-legible-tcn-epochs2000.json) |
| Normalized 2D joints + DTW | 50.77 ± 6.13% | 65.31 ± 4.54% | [JSON](fs-recognition-normalized-2d.json) |
| VIPE* + DTW | 64.35 ± 5.00% | 75.65 ± 3.62% | [JSON](fs-recognition-vipe.json) |
| VI-VPD + DTW | 65.92 ± 5.54% | 79.19 ± 2.92% | [JSON](fs-recognition-vi-vpd.json) |

The landing page imports the small [summary JSON](fs-recognition-summary.json). The raw files retain each support set, test prediction, source and feature hashes, runtime environment, and compact-model checkpoint hashes. All five downloaded result files matched their remote SHA-256 hashes.

## Shared evaluation

- Six jump types within figure skating: axel, lutz, flip, loop, salchow and toe loop.
- Official 2018 test performances: all 520 actions. The 183 validation labels are excluded from training and checkpoint selection.
- Released support manifests 0–4 for each shot count; the same support IDs and test labels across methods.
- Original action windows extend to cover at least 2.5 seconds before and 0.5 seconds after the jump midpoint. The supplied sequences contain 75–100 frames, median 81.
- Reference source pinned to [`7cde1cacd583162e39b7cd9f190818a76018b6dc`](https://github.com/jhong93/vpd/tree/7cde1cacd583162e39b7cd9f190818a76018b6dc). The adapter verifies source-file hashes and rejects predictions without a real selected reference example.

## What differs

**Compact classifier.** Each trial trains a fresh six-class LegibleTCN with 21,606 parameters and no pretraining. Its input uses the first released actor, 17 image-centered 2D joints, confidence, and position differences, resampled to 64 frames. Coordinates with confidence at most 0.01 are masked. Training uses AdamW, learning rate 0.002, weight decay 0.0001, label smoothing 0.1, cosine decay and no augmentation. The final epoch is evaluated. The 2,000-epoch arm starts separately rather than continuing the 80-epoch checkpoint. This is the benchmark classifier, distinct from the production live trajectory-template matcher.

**Reference matching.** The unmodified [`recognize.KnnModel('dtw', k=1)`](https://github.com/jhong93/vpd/blob/7cde1cacd583162e39b7cd9f190818a76018b6dc/recognize.py) uses Euclidean dynamic time warping, frame-vector normalization, variable-length sequences, and original/mirrored views. The official alternative DTW path behavior is retained; default-label substitution is rejected. Released offline gap interpolation is retained and can use later observations.

The [2D control](https://github.com/jhong93/vpd/blob/7cde1cacd583162e39b7cd9f190818a76018b6dc/dummy_2d_features.py) uses 13 joints, [hip centering and torso scaling](https://github.com/jhong93/vpd/blob/7cde1cacd583162e39b7cd9f190818a76018b6dc/vipe_dataset/dataset_base.py#L105-L139), and excludes confidence from its 26-value feature. VIPE* is a pretrained 64-value descriptor of 2D poses. VI-VPD is a 64-value descriptor learned from athlete crops and optical flow, with pose-and-motion supervision on additional unlabeled target-sport video. These runs consume released features; they do not train the descriptor from just 24 or 48 actions. The VPD frame extractor can use current and previous frames. Its published bidirectional detector and released offline preprocessing require separate treatment for a live application. See [VPD sections 3.1–3.3](https://jhong93.github.io/pdf/vpd-iccv21.pdf#page=3) and the [supplement](https://jhong93.github.io/pdf/vpd-supp.pdf).

## Interpretation

Longer compact-model training improves held-out accuracy, but transfer remains weak even though all ten 2,000-epoch trials classify 100% of their training examples correctly. The 80-epoch trials do not all fit their training examples. Extra training on the same support set does not measure the benefit of more varied labeled movement.

These whole-pipeline comparisons change normalization, confidence handling, representation, temporal alignment, views and prior training together. They do not isolate which change explains a particular share of the gap. Since the action cuts are supplied, this gap is not a segmentation penalty. The six fine-grained jump labels and clip accuracy also differ from [TOAD's 20 THUMOS14 action categories and per-frame mean average precision](https://arxiv.org/html/2501.13518v1#S4.SS3).

The [VPD project](https://jhong93.github.io/projects/vpd.html) publishes related results under its own experiment setup. The measured values above are retained as observed, including VI-VPD's 79.19% eight-example result, rather than replaced with a paper value.
