ajdanoff/asl-webcam-benchmark-augmented

ASL Webcam Benchmark — Augmented

An augmentation robustness benchmark for YOLO-based ASL digit and letter gesture detectors.

It contains 216 labeled webcam images: 36 original images and five augmented variants of each original image.

Intended use

Use this dataset to measure how gesture detectors respond to controlled changes in framing, scale, rotation, lighting, color, blur, sensor noise, and JPEG compression.

It is a stress-test benchmark, not a replacement for a larger multi-session held-out test set.

Dataset structure

data.yaml
augmentation_manifest.json
images/test/
labels/test/

Augmentations

Each original image has five deterministic augmented variants. Depending on the variant seed, an image may include:

• Horizontal flip
• Random crop
• Resize to 640 × 480
• Rotation
• Translation
• Scale change
• Brightness and contrast changes
• Gamma correction
• HSV hue and saturation changes
• Gaussian blur
• Gaussian noise
• JPEG compression

The exact transformations applied to every image are recorded in augmentation_manifest.json.

Class mapping

Class IDs	Symbols
0–9	0–9
10–35	A–Z

Evaluation protocol

Use symbol-based evaluation:

1. Convert each model prediction to a class symbol using that model’s own names mapping.
2. Convert each ground-truth class ID to a symbol using the shared 36-class mapping.
3. Count a true positive only when the symbols match and IoU is at least 0.45.

This protocol correctly supports:

• A unified model with class IDs 0–35
• A digit-only model with class IDs 0–9
• A letter-only model with class IDs 0–25

Known limitations

• All augmented variants derive from only one original image per class.
• Horizontal flips may produce visually ambiguous or unrealistic variants for some ASL gestures.
• Strong photometric and geometric transformations may be harder than the training distribution of some models.
• Results should be interpreted as augmentation robustness, not as a complete estimate of real-world gesture-recognition accuracy.