runrunrun1994 / Person_Segmentation

The inference implementation of the deeplabV3+ person segementation algorithm.

1. 文件目录说明

• Libtorch_PersonSegmentation Libtorch实现代码
• Ncnn_PersonSegmentation ncnn实现代码
• TensorRT_PersonSegmentation TensorRT实现代码
• ncnn andorid:https://github.com/runrunrun1994/ncnn-android-deeplabv3plus

2.实验结果

Backbone	Pixel Accuracy	Mean_IoU	Background IoU	Person IoU
Mobilenetv2_1.0	0.971	0.943	0.953	0.933
Mobilenetv2_0.5	0.966	0.933	0.944	0.922
Mobilenetv2_0.35	0.961	0.922	0.935	0.909
Mobilenetv2_0.25	0.955	0.911	0.926	0.896

Backbone	Params	FLOPS
Mobilenetv2_1.0	5.794M	23.260G
Mobilenetv2_0.5	3.282M	20.597G
Mobilenetv2_0.35	2.697M	19.993G
Mobilenetv2_0.25	2.344M	19.632G

• 数据分布、及组成待补充
• 速度对比待补充

Original	Result

3.引用

[1] Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
[2] Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation
[3] https://github.com/Tencent/ncnn
[4] https://github.com/pytorch/pytorch
[5] https://github.com/NVIDIA/TensorRT
其他引用待补充