Rui Peng
Rui Peng
I have limited ideas about the reason only from this succinct statement. Maybe there are something wrong on environment or fusion configuration?
我认为需要依据confidence的生成方式而定，回归方法（CVP）将每个像素点预测深度周围的四个深度假设层的权重之和作为confidence，而分类方法（R-MVSNet）或者我们的统一表示方法是将取所有假设层中的最大概率作为confidence。所以我们认为回归方法自然需要设置更大的prob_threshold来确保准确度。
That's it! We only experiment with FPN.
1. We adopt the PyTorch 1.2, CUDA 10.0. 2. Please ensure the fusion parameters (e.g. confidence threshold) are the same with us. 3. Maybe your fusible install failed due to...
可能有影响，没有做过相关实验，但训练和测试都用5个视图确实比训练用3个测试用5个要好。
Maybe you can try Pytorch1.2.0, cudatoolkit 10.0 and python3.6.
你的深度图是正常的吗？训练的时候深度图正常，测试不正常或者如果深度图是正常的，融合的点云是错误的可能是相机参数的问题（相机参数和图像尺寸不匹配也可能），或者点云是正常的但非常稀疏那可能是点云过滤阈值过大，比如光度一致性和几何一致性。
可能你的dataset加载训练样本有问题
This means that the object "self.color_aug" is a 'tuple', and I guess there are something wrong on the instantiation of "self.color_aug".
The test results are indeed related to the environment, maybe you can try **pytorch1.2**.