English | 简体中文
Currently FastDeploy using RKNPU2 to infer PPSeg supports the following model deployments:
| Model | Parameter File Size | Input Shape | mIoU | mIoU (flip) | mIoU (ms+flip) |
|---|---|---|---|---|---|
| Unet-cityscapes | 52MB | 1024x512 | 65.00% | 66.02% | 66.89% |
| PP-LiteSeg-T(STDC1)-cityscapes | 31MB | 1024x512 | 77.04% | 77.73% | 77.46% |
| PP-HumanSegV1-Lite(Universal portrait segmentation model) | 543KB | 192x192 | 86.2% | - | - |
| PP-HumanSegV2-Lite(Universal portrait segmentation model) | 12MB | 192x192 | 92.52% | - | - |
| PP-HumanSegV2-Mobile(Universal portrait segmentation model) | 29MB | 192x192 | 93.13% | - | - |
| PP-HumanSegV1-Server(Universal portrait segmentation model) | 103MB | 512x512 | 96.47% | - | - |
| Portait-PP-HumanSegV2_Lite(Portrait segmentation model) | 3.6M | 256x144 | 96.63% | - | - |
| FCN-HRNet-W18-cityscapes | 37MB | 1024x512 | 78.97% | 79.49% | 79.74% |
| Deeplabv3-ResNet101-OS8-cityscapes | 150MB | 1024x512 | 79.90% | 80.22% | 80.47% |
RKNPU needs to convert the Paddle model to RKNN model before deploying, the steps are as follows:
- For the conversion of Paddle dynamic diagram model to ONNX model, please refer to PaddleSeg Model Export.
- For the process of converting ONNX model to RKNN model, please refer to Conversion document.