Jia Guo

icon indicating an email [email protected]

icon location Hong Kong

Results 187 comments of Jia Guo

acc 一直是0

我这里用simplenet没问题，resnet需要加上se-block收敛更稳定

这个库考虑对一些具体的场景的 图片 做嘛

有标注数据的话应该是通用的

Pretrained model, detection and inference code

hi, maybe 2 months~

Pretrained model, detection and inference code

not sure yet.

Why number of labels has to be greater than 2 in MultiBoxTarget?

So we have to append '-1's if there are only 1 or 2 labels?

training speed

CPU is the performance bottleneck. We may need a refined c++ data loader to boost the speed.

Some parameters are confused for training wider face

It will end at epoch-80. Softmax already did per-instance normalization. Fixed parameters are correct. Color_jitter is required.

pretrained model requirst!!!

https://github.com/deepinsight/insightface/wiki/Model-Zoo

The image size is 640x640 or shorter side is 1200, long side is less than 1600?

Training with 640, but testing with any size you want.

The image size is 640x640 or shorter side is 1200, long side is less than 1600?

use default setting in config.py