download_cn.md 5.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
### 关键点检测模型
| 模型        | 输入尺寸 | AP (业务数据集) | AP (COCO Val) | 参数量 | FLOPS |单人推理耗时 (FP32) | 单人推理耗时(FP16) |             配置文件             |                           模型权重                           |                         预测部署模型                         |                  Paddle-Lite部署模型(FP32)                  |                  Paddle-Lite部署模型(FP16)                  |
| :---------- | :------: | :-----------: | :-----------: | :-----------: | :-----------: | :-----------------: | :-----------------: | :------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| PP-TinyPose |  128*96  |     84.3      |    58.4     | 1.32 M | 81.56 M | 4.57ms        |       3.27ms        | [Config](./tinypose_128x96.yml)  | [Model](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_128x96.pdparams) | [预测部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_128x96.zip) | [Lite部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_128x96_fp32.nb) | [Lite部署模型(FP16)](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_128x96_fp16.nb) |
| PP-TinyPose | 256*192  |     91.0      |   68.3       |  1.32 M  | 326.24M |14.07ms       |       8.33ms        | [Config](./tinypose_256x192.yml) | [Model](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_256x192.pdparams) | [预测部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_256x192.zip) | [Lite部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_256x192_fp32.nb) | [Lite部署模型(FP16)](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/tinypose_256x192_fp16.nb) |


### 行人检测模型
| 模型                 | 输入尺寸 | mAP (COCO Val-Person) | 参数量 | FLOPS | 平均推理耗时 (FP32) | 平均推理耗时 (FP16) |                           配置文件                           |                           模型权重                           |                         预测部署模型                         |                  Paddle-Lite部署模型(FP32)                  |                  Paddle-Lite部署模型(FP16)                  |
| :------------------- | :------: | :------------: | :------------: | :------------: | :-----------------: | :-----------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| PicoDet-S-Lcnet-Pedestrian | 192*192  |      31.7      | 1.16 M  | 170.03 M |  5.24ms        |       3.66ms        | [Config](../../picodet/application/pedestrian_detection/picodet_s_192_lcnet_pedestrian.yml) | [Model](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_192_lcnet_pedestrian.pdparams) | [预测部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_192_lcnet_pedestrian.zip) | [Lite部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_192_lcnet_pedestrian_fp32.nb) | [Lite部署模型(FP16)](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_192_lcnet_pedestrian_fp16.nb) |
| PicoDet-S-Lcnet-Pedestrian | 320*320  |      41.6      | 1.16 M   | 472.07 M | 13.87ms       |       8.94ms        | [Config](../../picodet/application/pedestrian_detection/picodet_s_320_lcnet_pedestrian.yml) | [Model](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_320_lcnet_pedestrian.pdparams) | [预测部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_320_lcnet_pedestrian.zip) | [Lite部署模型](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_320_lcnet_pedestrian_fp32.nb) | [Lite部署模型(FP16)](https://bj.bcebos.com/v1/paddledet/models/keypoint/tinypose_enhance/picodet_s_320_lcnet_pedestrian_fp16.nb) |

**说明**
- 关键点检测模型与行人检测模型均使用`COCO train2017`, `AI Challenger trainset`以及采集的多姿态场景数据集作为训练集。关键点检测模型使用多姿态场景数据集作为测试集,行人检测模型采用`COCO instances val2017`作为测试集。
- 关键点检测模型的精度指标所依赖的检测框为ground truth标注得到。
- 关键点检测模型与行人检测模型均在4卡环境下训练,若实际训练环境需要改变GPU数量或batch size, 须参考[FAQ](../../../docs/tutorials/FAQ/README.md)对应调整学习率。
- 推理速度测试环境为 Qualcomm Snapdragon 865,采用arm8下4线程推理得到。