Are you missing a convolution 3x3 layer in network?

[hcnhatnam]

After feature extraction layer(in you use Resnet50), paper used to a convolution 3x3 layer(3x3xC) to prepare lstm inputs.

[yizt]

@hcnhatnam 感谢您的反馈，确实是漏掉一个3x3卷积层，后面我会加上，再测试看看效果

[hcnhatnam]

@yizt Looking forward to your early feedback for result...

[yizt]

@hcnhatnam 增加了conv层后，效果没有明显区别

[hcnhatnam]

Can you update model again,please?

[yizt]

@hcnhatnam 抱歉，现已更新模型下载地址