Introduction
This article details the YuNet face detection model support in VisionPack and provides ready-to-use Deep View RT models.
Instructions
Performance
The following benchmarks were collected using the VisionPack detectgl application on an NXP i.MX 8M Plus EVK. The model is run on the NPU using the OV5640 camera as the capture device.
| Resolution | Framerate | Input Time | Decode Time | Inference Time |
| 320x240 | 110 FPS* | 1.5 ms | 0.75 ms | 6.5 ms |
| 400x300 | 75 FPS* | 2 ms | 1.5 ms | 9 ms |
| 480x360** | 50 FPS* | 2.5 ms | 2.5 ms | 16 ms |
| 512x384 | 60 FPS* | 2.5 ms | 3 ms | 12 ms |
*Framerate estimated in cases where model can run above display FPS. Camera is limited to 30 FPS but application will capture duplicate frames at display resolution which is 60 FPS.
**Model is per-channel quantized causing performance degradation.
Comments
0 comments
Please sign in to leave a comment.