关于小数据集训练

[aggestsfw]

感谢大佬分享这个很棒的工作，我用自己采集的数据，大约400张，用lidar做真值的，用blendedmvs_finetune.sh这个脚本去从头开始训练，大约30个epoch，可以收敛，但是预测不出来结果或者只有很少的点。 然后从BlendedMVS数据集中抽了其中一个场景来做同样的实验，也是预测不出来结果。 如果用整个BlendedMVS数据集来从头开始训练，不加载预训练模型，预测结果是正常的。 初始学习率设的是0.001。 DTU数据集部分的代码我也看了，就是同一个视角拍了七张不同打光的照片，其他基本上跟BlendedMVS数据集是一样的。 百思不得其解，想请教下大佬为什么小数据集训练会有问题？

[prstrive]

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。

[aggestsfw]

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。

谢谢你的答复，这是我刚刚跑的，用的是BlendedMVS里，BlendedMVS/dataset_low_res/58cf4771d0f5fb221defe6da这个场景，只有64张照片，训练一开始Th2、Th4、Th8输出都是0，预测出来没有点云

双gpu，batchsize=1 输出是这样的：

Epoch 0/16: 100% 32/32 || Elapsed Time: 0:01:02,Time: 0:01:02,LR: 0.001,Loss: 0.223|0.414,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.253|0.248,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 1/16: 100% 32/32 || Elapsed Time: 0:00:44,Time: 0:00:44,LR: 0.001,Loss: 0.282|0.242,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.297|0.286,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 2/16: 100% 32/32 || Elapsed Time: 0:00:42,Time: 0:00:42,LR: 0.001,Loss: 0.335|0.324,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:08,Time: 0:00:08,Loss: 0.362|0.349,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 3/16: 15% 05/32 || Elapsed Time: 0:00:09,ETA: 0:00:51,LR: 0.001,Loss: 0.352|0.350,Th2Epoch 3/16: 18% 06/32 || Elapsed Time: 0:00:10,ETA: 0:00:47,LR: 0.001,Loss: 0.356|0.351,Th2Epoch 3/16: 21% 07/32 || Elapsed Time: 0:0EpEpoch 3/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.001,Loss: 0.340|0.353,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000,Loss: 0.350|0.352,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.345|0.351,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 4/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.001,Loss: 0.334|0.341,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.310|0.316,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 5/16: 100% 32/32 || Elapsed Time: 0:00:42,Time: 0:00:42,LR: 0.001,Loss: 0.304|0.321,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:08,Time: 0:00:08,Loss: 0.302|0.310,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 6/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.001,Loss: 0.289|0.302,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.287|0.294,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 7/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.001,Loss: 0.294|0.287,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.272|0.281,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 8/16: 100% 32/32 || Elapsed Time: 0:00:42,Time: 0:00:42,LR: 0.001,Loss: 0.295|0.280,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:08,Time: 0:00:08,Loss: 0.269|0.275,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 9/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.001,Loss: 0.271|0.275,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.265|0.275,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 10/16: 100% 32/32 || Elapsed Time: 0:00:43,Time: 0:00:43,LR: 0.0005,Loss: 0.267|0.272,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Val:100% 32/32 || Elapsed Time: 0:00:09,Time: 0:00:09,Loss: 0.263|0.272,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.000 Epoch 11/16: 43% 14/32 || Elapsed Time: 0:00:20,ETA: 0:00:26,LR: 0.0005,Loss: 0.260|0.270,Th2: 0.000|0.000,Th4: 0.000|0.000,Th8: 0.000|0.0

[prstrive]

可能你的dataset加载训练样本有问题

[aggestsfw]

可能你的dataset加载训练样本有问题

谢谢，已经解决了，自己的数据是训练时间不够