tongfeiliu / c3net-for-building-extraction

• 👋 Hi, I’m @TongfeiLiu
• 👀 I’m interested in AI-based remote sensing image understanding.
• 🌱 I’m currently a Lecturer at the School of Electronic Information and Artificial Intelligence of Shaanxi University of Science & Technology.
• 💞️ I’m looking to collaborate on ...
• 📫 Welcome to communicate with me by email ([email protected])

c3net-for-building-extraction's Issues

训练时bce_loss计算错误

训练时，在bce_Loss = BCE_loss(pred1, label)报错，查看之后label是([2, 1, 256, 256])，pred1是([2, 2, 256, 256])，请问怎么解决这个问题啊

Some confusion expecting a reply

Great work!

I'm a beginner to deep learning. I want to make some changes on your work to improve segmentation accuracy, but something confused me when running the project.

1. For the Inria dataset preparation, the images are 5000×5000, and for how to crop the non-overlapping clipping to 81 sheets in 512 ×512. Is the image I thought below correct？
   

2. I train 100 epoches, and save the 100th epoch model. The mIoU looks higer and I don't know the reason. (The grey is my test results.)
   

3. I used another dataset. It has a small number of pictures. When performing the ablation experiment, I annotated C2AM and ER2M in the network construction and only used BCE Loss, but the best mIoU result in 100 rounds of training were also achieved and even better. I don't know what's wrong here.
   code:
   
   C3Net-BD:
   
   C3Net-ablation-B:
   

4. Comparing the article and the code, this one feels inconsistent, but the experimental results feel not much different.
   
   

It may be that I didn't look carefully in some places. Looking forward to your reply and hope you can solve my confusion.