ajdanoff/asl-webcam-benchmark-original

ASL Webcam Benchmark — Original

A small held-out benchmark for evaluating YOLO-based ASL digit and letter gesture detectors on original webcam frames.

The dataset contains one original webcam image for each of 36 gesture classes: digits 0–9 and letters A–Z.

Intended use

This benchmark is intended as a quick single-frame robustness check. Because it contains only one image per class, it should not be used as the sole basis for model selection or statistical comparison.

Dataset structure

data.yaml
images/test/
labels/test/

Each image has one YOLO label file containing one normalized bounding box:

class_id x_center y_center width height

Class mapping

Class IDs	Symbols
0–9	0–9
10–35	A–Z

The class order is:

0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
A, B, C, D, E, F, G, H, I, J,
K, L, M, N, O, P, Q, R, S, T,
U, V, W, X, Y, Z

Evaluation protocol

A prediction is counted as correct when:

1. Its predicted class symbol matches the ground-truth symbol.
2. Its predicted bounding box has IoU of at least 0.45 with the ground-truth box.
3. Its confidence is at least the evaluation threshold.

The benchmark uses symbol-based evaluation so that specialist digit and letter models can be compared with a unified 36-class model despite using different internal class-index ranges.

Known limitations

• Contains only one image per gesture class.
• Results are highly sensitive to the selected frame.
• It is not suitable for calculating stable per-class confidence intervals.
• It should be used together with the augmented benchmark and a larger multi-frame held-out set.

Provenance

Images were captured with webcams and converted to YOLO format using MediaPipe hand-landmark boxes. The conversion manifest records the MediaPipe, YOLO fallback, Kalman fallback, and centered-fallback box sources.