资讯 社区 文档
技术能力
语音技术
文字识别
人脸与人体
图像技术
语言与知识
视频技术

EdgeBoard支持的模型

目前,EdgeBoard主要支持视觉类模型,包含分类、检测、分割以及特定场景模型。

分类模型

模型为PaddleClas release/static的ModelZoo,github地址:https://github.com/PaddlePaddle/PaddleClas/tree/release/static

以下数据的测试环境:

模型训练环境:PaddlePaddle = 1.8.5、PaddleClas = release/static

模型部署环境:EdgeBoard 软核1.8.1、系统Ubuntu18.04

ResNet and Vd series

ResNet model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3A time by 3A(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
ResNet18 0.7098 0.8992 0.7096 0.8991 14.71548588 0.733 0.914 8.95363345 0.7097 0.8988 8.95363345 0.7097 0.8988 6.40590735
ResNet18_vd 0.7226 0.908 0.723 0.908 16.02208967 0.7223 0.9081 9.83267931 0.7223 0.9081 17.8796712 0.7221 0.908 7.09548706
ResNet34 0.7457 0.9214 0.7441 0.9211 26.98735892 0.7441 0.921 15.6827087 0.7441 0.921 15.6827087 0.7437 0.9211 10.6187022
ResNet34_vd 0.7598 0.9298 0.7586 0.9303 28.36009001 0.7582 0.93 16.5780194 0.7582 0.93 16.5780194 0.7587 0.9298 11.3229563
ResNet50 0.765 0.93 0.7637 0.9294 34.05404204 0.7641 0.9294 21.732892 0.7641 0.9294 21.732892 0.7639 0.9296 16.6979658
ResNet50_vc 0.7835 0.9403 0.7827 0.9398 35.21767119 0.7903 0.9446 22.9020555 0.7903 0.9446 22.9020555 0.7823 0.9398 17.2306128
ResNet50_vd 0.7912 0.9444 0.7904 0.9446 35.71911547 0.7741 0.9365 37.4155384 0.7741 0.9365 37.4155384 0.7903 0.9444 17.6102096
ResNet101 0.7756 0.9364 0.7742 0.9366 59.52945458 0.8013 0.9493 38.5317895 0.8013 0.9493 38.5317895 0.7742 0.9361 27.3622326
ResNet101_vd 0.8017 0.9497 0.8014 0.9495 61.17122603 0.7823 0.939 53.0845316 0.7823 0.939 53.0845316 0.8014 0.9491 28.3110359
ResNet152 0.7826 0.9396 0.7827 0.9393 85.05916077 0.8041 0.9523 54.2026246 0.8041 0.9523 54.2026246 0.782 0.939 37.6998016
ResNet152_vd 0.8059 0.953 0.8041 0.9523 86.59876956 0.8075 0.9534 68.9850622 0.8075 0.9534 68.9850622 0.8043 0.9522 38.5713139
ResNet200_vd 0.8093 0.9533 0.807 0.9537 110.5979351 0.8069 0.9538 49.9698168 0.8069 0.9538 49.9698168 0.8069 0.9538 49.9698168

Mobile series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top1 Acc by 5C Top-5 Acc by 5C time by 5C(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
MobileNetV1_x0_25 0.5143 0.7546 0.5025 0.7443 3.68112385 0.5025 0.7443 3.47995238 0.5025 0.7443 4.05508755 0.5025 0.7443 4.05508755 0.5025 0.7443 3.47995238
MobileNetV1_x0_5 0.6352 0.8473 0.6325 0.8469 4.97116533 0.6325 0.8469 4.35903775 0.6325 0.8469 5.05141552 0.6325 0.8469 5.05141552 0.6325 0.8469 4.35903774
MobileNetV1_x0_75 0.6881 0.8823 0.6838 0.8803 7.60554075 0.6843 0.8802 6.29608946 0.6843 0.8802 7.24652382 0.6843 0.8802 7.24652382 0.6843 0.8802 6.29608946
MobileNetV1 0.7099 0.8968 0.7079 0.8956 9.34210897 0.708 0.8954 7.22738171 0.708 0.8954 8.4207836 0.708 0.8954 8.4207836 0.708 0.8954 7.05951087
MobileNetV2_x0_25 0.5321 0.7652 0.5273 0.7616 7.74662709 0.5269 0.7616 7.09189643 0.5273 0.7617 8.24141499 0.5273 0.7617 8.24141499 0.5269 0.7616 6.96900913
MobileNetV2_x0_5 0.6503 0.8572 0.6453 0.8541 7.4191766 0.5135 0.6808 6.86391964 0.6451 0.8539 8.13312565 0.6451 0.8539 8.13312565 0.5936 0.7877 6.75820331
MobileNetV2_x0_75 0.6983 0.8901 0.6959 0.8879 12.7507112 0.6959 0.8878 11.5829853 0.696 0.8878 13.0029739 0.696 0.8878 13.0029739 0.6959 0.8878 11.3544805
MobileNetV2 0.7215 0.9065 0.7186 0.9052 11.0632651 0.7188 0.9051 9.86122213 0.7188 0.9052 11.3809789 0.7188 0.9052 11.3809789 0.7188 0.9051 9.66140689
MobileNetV2_x1_5 0.7412 0.9167 0.7385 0.9155 17.1437208 0.7345 0.9133 14.5713013 0.7345 0.9133 16.5544266 0.7345 0.9133 16.5544266 0.8014 0.9491 14.2934955
MobileNetV2_x2_0 0.7523 0.9258 0.7386 0.9154 22.0435947 0.3747 0.6042 17.6118443 0.7384 0.9153 20.1993121 0.7384 0.9153 20.1993121 0.7826 0.9397 17.1324282
GhostNet_x1_0 0.7402 0.9165 0.4417 0.677 34.1087561 0.5221 0.7592 30.9889381 0.3239 0.5406 33.5847206 0.3239 0.5406 33.5847206 0.5221 0.7592 30.9889381

SEResNeXt and Res2Net series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
Res2Net50_26w_4s 0.7933 0.9457 0.7923 0.945 46.9715829 0.7923 0.945 41.3515309
Res2Net50_vd_26w_4s 0.7975 0.9491 0.797 0.9484 48.1105378 0.7969 0.9485 42.0706068
Res2Net50_14w_8s 0.7946 0.947 0.7875 0.9414 64.7512679 0.7873 0.9412 57.5119234
Res2Net101_vd_26w_4s 0.8064 0.9522 0.7938 0.9393 74.3632303 0.7933 0.9394 62.2968839
Res2Net200_vd_26w_4s 0.8121 0.9571 0.7967 0.9418 132.182714 0.7967 0.9413 108.708479
ResNeXt50_32x4d 0.7775 0.9382 0.7757 0.937 27.6730268 0.7675 0.9281 22.2482438
ResNeXt50_vd_32x4d 0.7956 0.9462 0.7757 0.937 27.6730268 0.74 0.96 23.0841588
ResNeXt50_64x4d 0.7843 0.9413 0.7836 0.9403 44.1621114 0.7843 0.9413 52.8567324
ResNeXt50_vd_64x4d 0.8012 0.9486 0.8002 0.9483 45.2178389 0.8012 0.9486 54.1873319
ResNeXt101_32x4d 0.7865 0.9419 0.7854 0.9415 44.4215912 0.7865 0.9419 33.8506747
ResNeXt101_vd_32x4d 0.8033 0.9512 0.8013 0.9501 45.5457859 0.8033 0.9512 34.8075458
ResNeXt101_64x4d 0.7835 0.9452 0.7925 0.9452 72.0490677 0.792 0.9449 82.8039959
ResNeXt101_vd_64x4d 0.8078 0.952 0.8027 0.9472 73.287074 0.8078 0.952 81.9382522
ResNeXt152_32x4d 0.7898 0.9433 0.7888 0.943 62.0476544 0.7889 0.9432 48.0175154
ResNeXt152_vd_32x4d 0.8072 0.952 0.8063 0.9519 63.1244978 0.8072 0.952 47.7385326
ResNeXt152_64x4d 0.7951 0.9471 0.7937 0.946 101.227566 0.7951 0.9471 113.319645
ResNeXt152_vd_64x4d 0.8108 0.9534 0.8102 0.9532 102.414722 0.8107 0.9533 113.768651
SE_ResNet18_vd 0.7333 0.9138 0.733 0.9142 12.8133037 0.7334 0.913 9.29966265
SE_ResNet34_vd 0.7651 0.932 0.7641 0.9313 22.3687792 0.7651 0.932 15.6554105
SE_ResNet50_vd 0.7952 0.9475 0.7946 0.9473 34.68155 0.7952 0.9475 27.4156064
SE_ResNeXt50_32x4d 0.7844 0.9396 0.7832 0.9396 39.3880242 0.7844 0.9396 32.3298221
SE_ResNeXt50_vd_32x4d 0.8024 0.9489 0.8024 0.9489 42.8019707 0.7922 0.9386 34.4062971
SE_ResNeXt101_32x4d 0.7912 0.942 0.7924 0.9438 66.9515808 0.7834 0.942 53.2186031
SENet154_vd 0.814 0.9548 0.8127 0.9542 138.551309 0.814 0.9548 143.699729

DPN and DenseNet series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3A time by 3A(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
DenseNet121 0.7566 0.9258 0.7558 0.9258 53.6027625 0.7558 0.9258 53.6027625 0.7558 0.9259 44.3492407 0.7566 0.9258 39/3477994
DenseNet161 0.7857 0.9414 0.7844 0.9406 115.46785 0.7844 0.9406 115.46785 0.7845 0.9411 91.7863596 0.7857 0.9414 80.0224323
DenseNet169 0.7681 0.9331 0.768 0.9328 68.6313306 0.768 0.9328 68.6313306 0.7675 0.9328 57.4114082 0.7681 0.9331 50.6865443
DenseNet201 0.7763 0.9366 0.7755 0.9364 90.441223 0.7755 0.9364 90.441223 0.7757 0.9366 75.0129829 0.7763 0.9366 66.8523895
DenseNet264 0.7796 0.9385 0.7793 0.9388 129.401795 0.7793 0.9388 129.401795 0.779 0.9389 105.48438 0.7796 0.9385 93.7701779
DPN68 0.7678 0.9343 0.724 0.9092 55.3380497 0.724 0.9092 55.3380497 0.7243 0.9093 44.8887849 0.7678 0.9343 40.6067176
DPN92 0.7985 0.948 0.7979 0.9478 132.163013 0.7979 0.9478 132.163013 0.7983 0.9478 96.3646919 0.7985 0.948 85.7311024
DPN98 0.8059 0.951 0.803 0.9496 166.167311 0.803 0.9496 166.167311 0.8035 0.9501 116.329713 0.8059 0.951 121.238888
DPN107 0.8089 0.9532 0.7019 0.8887 254.276414 0.7019 0.8887 254.276414 0.7015 0.8878 174.463986 0.8089 0.9532 181.168191
DPN131 0.807 0.9514 0.8034 0.9489 223.894086 0.8034 0.9489 223.894086 0.8036 0.949 157.157379 0.807 0.9514 163.428369

HRNet series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3A time by 3A(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
HRNet_W18_C 0.7692 0.9339 0.5661 0.8185 60.9819089 0.5661 0.8185 60.9819089 0.5659 0.8186 55.4794437 0.7873 0.9412 39.3477994
HRNet_W30_C 0.7804 0.9402 0.657 0.8775 83.3555966 0.657 0.8775 83.3555966 0.6571 0.8782 65.7723425 0.7969 0.9485 44.0211977
HRNet_W32_C 0.7828 0.9424 0.7821 0.9425 82.9289218 0.7821 0.9425 82.9289218 0.7821 0.9423 64.4770416 0.7923 0.945 50.0433493
HRNet_W40_C 0.7877 0.9447 0.7839 0.9428 111.186797 0.7839 0.9428 111.186797 0.7835 0.9432 80.2612852 0.7933 0.9394 60.8345537
HRNet_W44_C 0.79 0.9451 0.766 0.9349 126.927966 0.766 0.9349 126.927966 0.7663 0.9349 88.4218833 0.7967 0.9413 65.3390154
HRNet_W48_C 0.7895 0.9442 0.7892 0.9443 135.097451 0.7892 0.9443 135.097451 0.7894 0.9441 90.5619071 0.7936 0.9677 65.9286927
HRNet_W64_C 0.793 0.9461 0.7926 0.9462 206.795211 0.7926 0.9462 206.795211 0.7929 0.946 122.473623 0.7959 0.9478 83.8883701

Inception series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3A time by 3A(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
Xception41_deeplab 0.7955 0.9438 0.7593 0.9214 68.4773261 0.7593 0.9214 68.4773261 0.7593 0.9215 47.2026896 0.7923 0.945 38.5318629
Xception65_deeplab 0.8032 0.9449 0.7735 0.929 103.461153 0.7735 0.929 103.461153 0.7743 0.9287 69.1011319 0.7967 0.9413 55.9068845
InceptionV3 0.7914 0.9459 0.7348 0.9108 27.4721269 0.7348 0.9108 27.4721269 0.735 0.9104 21.0058066 0.6959 0.8878 16.2281211
InceptionV4 0.8077 0.9526 0.7591 0.9218 54.4613488 0.7591 0.9218 54.4613488 0.7591 0.9214 38.6780565 0.7188 0.9051 29.8487604

EfficientNet and ResNeXt101_wsl series

Model Top-1 Acc by PaddleClas Top-5 Acc by PaddleClas Top-1 Acc by 3A Top-5 Acc by 3A time by 3A(ms) Top-1 Acc by 3B Top-5 Acc by 3B time by 3B(ms) Top-1 Acc by 5D Top-5 Acc by 5D time by 5D(ms) Top-1 Acc by 9D Top-5 Acc by 9D time by 9D(ms)
ResNeXt101_32x8d_wsl 0.8255 0.9674 0.8223 0.9659 133.606047 0.8223 0.9659 133.606047 0.8225 0.9656 67.4824983 0.82 0.95 47.6001574
ResNeXt101_32x16d_wsl 0.8424 0.9726 0.8385 0.9712 83.7354768 0.8385 0.9712 83.7354768 0.8385 0.9712 151.637222 0.8312 0.9714 82.075878
ResNeSt50_fast_1s1x64d 0.8035 0.9528 0.8022 0.9521 40.8778803 0.8022 0.9521 40.8778803 0.802 0.952 28.9425293 0.7973 0.9612 21.9779988

检测模型

模型为paddledetection release/0.5的ModelZoo,github地址:https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.5/docs/MODEL_ZOO_cn.md

以下数据的测试环境:

模型训练环境:PaddlePaddle = 1.8.5、PaddleDetection = release/0.5

模型部署环境:EdgeBoard 软核1.8.1、系统Ubuntu

Yolov3 on pascal voc

骨架网络 输入尺寸 mAP by paddledetection mAP on FZ9D
DarkNet53 608 83.5 81.6
DarkNet53 416 83.6 80.9
DarkNet53 320 82.2 81.2
DarkNet53 Diou-Loss 608 83.5 83.3
MobileNet-V1 608 76.2 75.6
MobileNet-V1 416 76.7 74.3
MobileNet-V1 320 75.3 73.1
ResNet34 608 82.6 80.2
ResNet34 416 81.9 81.1
ResNet34 320 80.1 78.2

Yolov3 on COCO

骨架网络 输入尺寸 加入deformable卷积 mAP by paddledetection mAP by FZ9D mAP by FZ5C FZ5C time cost
DarkNet53 608 38.9 38.7 38.5 248.9
DarkNet53 416 37.5 37.2 34.5 83
DarkNet53 320 34.8 33.5 34.5 82.95
MobileNet-V1 608 29.3 28.7 28.9 155
MobileNet-V1 416 29.3 28.8 26.7 50
MobileNet-V1 320 27.1 26.9 27.6 50
ResNet34 608 36.2 36 35.86 177.4
ResNet34 416 34.3 33.9 31 59.3
ResNet34 320 31.4 30.5 31.1 59

SSD on pascal voc

骨架网络 输入尺寸 mAP by paddledetection mAP on EdgeBoard
MobileNet v1 300 73.2 暂无数据

其他支持的模型

分类 模型 支持平台
分割模型 deeplabv3P FZ3、FZ5、FZ9
关键点 HRnet FZ3、FZ5、FZ9
特定场景 人体检测 场景化软硬一体设备
特定场景 人体关键点检测 场景化软硬一体设备
特定场景 人脸识别 场景化软硬一体设备
特定场景 安全帽佩戴合规检测 场景化软硬一体设备
特定场景 烟火检测 场景化软硬一体设备
特定场景 车辆检测 场景化软硬一体设备
特定场景 车型识别 场景化软硬一体设备
特定场景 动态车流 场景化软硬一体设备
特定场景 车辆属性 场景化软硬一体设备
特定场景 车牌识别 场景化软硬一体设备

模型库持续更新中...

上一篇
EdgeBoard支持的算子
下一篇
硬件说明