A human gesture recognition model for the American Sign Language (ASL) recognition scenario (word-level recognition). The model uses an S3D framework with MobileNet V3 backbone. Please refer to the MS-ASL-100* dataset specification to see the list of gestures that are recognized by this model.
The model accepts a stack of frames sampled with a constant framerate (15 FPS) and produces a prediction on the input clip.
| Metric | Value |
|---|---|
| Top-1 accuracy (MS-ASL-100*) | 0.847 |
| GFlops | 6.660 |
| MParams | 4.133 |
| Source framework | PyTorch* |
Name: input, shape: [1x3x16x224x224]. An input image sequence in the format [BxCxTxHxW], where:
- B - batch size
- C - number of channels
- T - duration of input clip
- H - image height
- W - image width
The model outputs a tensor with the shape [Bx100], each row is a logits vector of performed ASL gestures.
[*] Other names and brands may be claimed as the property of others.