DRIVE数据集效果较差疑问

[favoMJ]

我尝试在DRIVE上进行训练，epoch设置为200，batch_size设置为512；在train.sh进行训练，test.sh生成pred，再通过eval将生成的pred和label得到的结果比较差，能帮忙分析下原因吗

[xiongxyowo]

Generally , fundus images place greater emphasis on the preservation of local details, so simply increasing the size of the encoder offers limited benefits. You may refer to our work [*1] or explore methods to construct custom non-pretrained/patching encoders (which is currently the mainstream strategy for fundus image segmentation), paired with a full-resolution decoder. The final output of SAM2-UNet is obtained through 4× upsampling, making it unsuitable for tasks like retinal image analysis that heavily rely on fine details.
[*1] https://ieeexplore.ieee.org/document/11107345

一般来说，眼底图像一般更侧重于研究对细节的保留，因此提升编码器的容量贡献是有限的。你可以参考我们的工作[*1]，或者研究如何构建自定义的非预训练/切片编码器，这是目前眼底图像分割的主流策略。SAM2-UNet的最终结果通过四倍上采样得到，不适用于眼底图像这种极其依赖细节的任务。

[Yu-zhengbo]

Even the model didn't converge in my research datasets, I think it's influenced by the target of the task.