xiakj

@liutianyuan 我的也是，模糊图像和底库相似度也挺高。也不知道该怎么解决，目前是把低置信度的人脸框、侧脸过滤掉不处理了，来降低一点风险。

@mzxhzhp 什么意思，底库不就是图像的特征值么

@Jasonmes 怎么解决的

@107318095 @Challovactor where the walker feature extraction model?

@wzhiyuan2016 看起来是backbone设计的太简单了，特征少了

Hi, have you solved this problem? How?

similar bug. mark

log file : usage: main.py [-h] [--net_type NET_TYPE] [--queue_size QUEUE_SIZE] [--print_freq PRINT_FREQ] [--pretrained_model_path PRETRAINED_MODEL_PATH] [--batch_size BATCH_SIZE] [--alpha ALPHA] [--loss_type {Arc,AM,SV}] [--margin MARGIN] [--scale SCALE] [--neg_margin NEG_MARGIN] [--sync_bn] [--feat_dim FEAT_DIM] ip...

@tiandunx yes, I set "world_size" 8, then the mistaken happened

@tiandunx my train_ffc.sh file is: ip=127.0.0.1 # your ip address port= # int port number export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 world_size=8 exe=main.py queue_size=851 # #total_identities * ratio net_type=mobile # if 224 x 224,...