Mobile Detection Benchmark
November 3, 2021 ยท View on GitHub
This repo is used to test the speed of the mobile terminal models
Benchmark Result
| Model | Input size | mAPval 0.5:0.95 | mAPval 0.5 | Params (M) | FLOPS (G) | Latency1 (ms) | Latency2 (ms) | Config |
|---|---|---|---|---|---|---|---|---|
| YOLOv3-Tiny | 416 | 16.6 | 33.1 | 8.86 | 5.62 | 25.42 | - | model link |
| YOLOv4-Tiny | 416 | 21.7 | 40.2 | 6.06 | 6.96 | 23.69 | - | model link |
| PP-YOLO-Tiny | 320 | 20.6 | - | 1.08 | 0.58 | 6.75 | - | model link |
| PP-YOLO-Tiny | 416 | 22.7 | - | 1.08 | 1.02 | 10.48 | - | model link |
| Nanodet-M | 320 | 20.6 | - | 0.95 | 0.72 | 8.71 | - | model link |
| Nanodet-M | 416 | 23.5 | - | 0.95 | 1.2 | 13.35 | - | model link |
| Nanodet-M 1.5x | 416 | 26.8 | - | 2.08 | 2.42 | 15.83 | - | model link |
| YOLOX-Nano | 416 | 25.8 | - | 0.91 | 1.08 | 19.23 | - | model link |
| YOLOX-Tiny | 416 | 32.8 | - | 5.06 | 6.45 | 32.77 | - | model link |
| YOLOv5n | 640 | 28.4 | 46.0 | 1.9 | 4.5 | 40.35 | - | model link |
| YOLOv5s | 640 | 37.2 | 56.0 | 7.2 | 16.5 | 78.05 | - | model link |
| PicoDet-S | 320 | 27.1 | 41.4 | 0.99 | 0.73 | 8.13 | 6.65 | model link |
| PicoDet-S | 416 | 30.6 | 45.5 | 0.99 | 1.24 | 12.37 | 9.82 | model link |
| PicoDet-M | 320 | 30.9 | 45.7 | 2.15 | 1.48 | 11.27 | 9.61 | model link |
| PicoDet-M | 416 | 34.3 | 49.8 | 2.15 | 2.50 | 17.39 | 15.88 | model link |
| PicoDet-L | 320 | 32.6 | 47.9 | 3.24 | 2.18 | 15.26 | 13.42 | model link |
| PicoDet-L | 416 | 35.9 | 51.7 | 3.24 | 3.69 | 23.36 | 21.85 | model link |
| PicoDet-L | 640 | 40.3 | 57.1 | 3.24 | 8.74 | 54.11 | 50.55 | model link |
| PicoDet-Shufflenetv2 1x | 416 | 30.0 | 44.6 | 1.17 | 1.53 | 15.06 | 10.63 | model link |
| PicoDet-MobileNetv3-large 1x | 416 | 35.6 | 52.0 | 3.55 | 2.80 | 20.71 | 17.88 | model link |
| PicoDet-LCNet 1.5x | 416 | 36.3 | 52.2 | 3.10 | 3.85 | 21.29 | 20.8 | model link |
Table Notes:
- Latency: All our models test on
Qualcomm Snapdragon 865(4\*A77+4\*A55)with 4 threads by arm8 and with FP16. In the above table, test latency on1NCNN and2Paddle-Lite. - All model are trained on COCO train2017 dataset and evaluated on COCO val2017.