A reconstruction can look plausible and still leave the research question unanswered. For KineticAugment, the important question is whether plausible body and hand transformations can produce useful training examples that improve sign recognition. A round-trip assessment checks an earlier link in that chain: whether landmarks survive conversion through body and hand models. It does not test a recognizer.
The sample and measurements
On 7 October 2026, the assessment drew 20 random non-pilot clips from a pool of 329 WLASL clips. The clips contained 1,236 frames total and ranged from 288×192 to 1920×1080 pixels, with 24 to 106 frames per clip. The script measured MediaPipe → SMPL-X → MediaPipe body landmarks and MediaPipe → MANO → MediaPipe hand landmarks.
The pooled body direct error was 3.13% of shoulder width. For hands, the in-sample MANO mean was 5.10 mm and the hand detector found a hand in 42% of side-frames. These units matter: body error is normalized by shoulder width, while hand error is millimeters. Neither is recognition accuracy.
The five-clip pilot had pooled direct body error of 3.52% and pooled held-out body error of 3.99%; the random 20 had 3.13% and 3.37%, respectively. For hands, the pilot in-sample mean was 7.38 mm compared with 5.10 mm for the 20-clip sample. On the assessment’s provisional targets, hand tips and the 95th percentile passed, while mean error missed the 5 mm target by 0.1 mm.
What the result supports
Within this sample, the body round trip behaved similarly to the pilot by median error, while the hand mean was lower. The body direct error varied from 1.44% to 5.52% across clips. A single pooled score hides that spread and the fact that some errors grew when an elbow bent or a hand rose toward the face.
The sample is useful as a transfer check within the same source pool. It is not evidence of cross-dataset generalization: some clips share a signer, and the two clips 08917 and 08918 show the same signer performing the same sign twice. Hand coverage is also uneven. Only six of the 20 clips had a scored left hand, so the hand summary largely describes dominant-hand tracks. The hand detector coverage of 42% of side-frames is itself a practical limitation.
These measurements assess a round trip from detected landmarks through fitted geometry and back. They do not say whether a transformed clip looks natural to signers, preserves linguistic meaning, or provides useful variation. They also do not show that a recognizer trained with generated examples improves.
What a useful augmentation test needs
The augmentation question requires a controlled recognizer experiment. First partition original videos into training, validation, and test sets. Generate candidate versions only for training clips and keep each generated version with its source in the training partition. Fit the same recognizer with the same preprocessing and training budget on (1) original training clips and (2) original clips plus accepted augmentations.
Use untouched original validation clips for model selection and untouched original test clips for the final comparison. Where metadata permits, audit signer and source overlap across splits. If augmented versions of validation clips are examined, report that as a separate robustness view: they do not increase the number of independent validation examples.
Acceptance rules for candidate augmentations should be based on checks that can catch pose-dependent errors, detector failures, and temporal defects. A low average round-trip error alone is not enough. The recognizer comparison is still the decisive test.
Source and further reading
The figures and limitations above come from the public Week 2 findings, dated 7 October 2026. The report was calculated from WLASL footage outside the repository; no footage is reproduced here. See the KineticAugment research page for the project question and evaluation plan.