Shao-Yuan Lo
Shao-Yuan Lo
The test set mIoU 67.3 was obtained by two-stage training. For the val set, we only did one-stage training. We got mIoU 65.1. You can find detailed information in our...
The models for ablation study (Table 1) are trained by 150 epochs. The model for the final evaluation (Table 2) is trained by two-stage. The first stage has 150 epochs,...
Hi @CS-Jackson Yes, your statement is correct. But please note that the first stage is trained by downsampled annotations (1/8 resolution, i.e., do not upsample the feature maps), and the...
@CS-Jackson for the training sample (512,1024)
My Titan X is "Titan X Maxwell"! So it seems reasonable that the speed on my Titan X is significantly lower than GTX 1080Ti. Thanks for your reminder. :)
DenseNet also has several Dense Block, and there is a transition layer between every two adjacent blocks. In my work, EDA Blocks can be corresponding to Dense Blocks, and Downsampling...
Hi, thanks for your interest in our paper! Unfortunately, as I stated in README, I cannot share the source code now due to the sponsor's request. But I still hope...