search

cuiziteng / Aleth-NeRF

🌕 [AAAI 2024] Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (Low-light enhance / Exposure correction + NeRF)
Apache License 2.0
76 stars 5 forks source link

如何自制数据集 #19

Closed XLR-man closed 4 months ago

XLR-man commented 5 months ago

请问我自己拍的数据集，怎么得到transform_test.json这些分开的json文件？ 先通过colmap，得到稀疏数据，怎么得到你数据集中的其他json文件呢？

cuiziteng commented 4 months ago

你好，我是通过torch-ngp项目中的code生成的json文件，你可以参考 https://github.com/ashawkey/torch-ngp

image

至于如何划分训练测试自己定义就好

XLR-man commented 4 months ago

好的感谢。 另外还有一个问题，因为你的模型是训练低光场景的，那我如果我想输入正常光场景进行训练能成功吗？我明白你的任务，我是想输入正常光场景进行训练，看一下一些中间变量与低光的有什么区别。

XLR-man commented 4 months ago

我输入正常光场景进行训练时，发现loss都为nan值，进一步查看发现重建出来的颜色值非常低，类似低光场景的那种低，但是这时gt是正常光的值，这样一做损失就为nan值了。请问为什么输入正常光场景进行训练无法训练？这是有什么原因吗？ 我所理解的是，一个场景下的正常光和低光只是亮度不同，位姿等坐标信息都是一样的，那经过NeRF框架之后，我与正常光的值做损失，那也应该能训练出正常光的NeRF吧？

cuiziteng commented 4 months ago

训练正常光照的时候建议把Inverse Tone MSE loss换回MSE loss。

6.7更： 之前Inverse Tone MSE那里忘记加CLIP了，现在加了一下，可以使用了。

lucky9-cyou commented 4 months ago

训练正常光照的时候建议把Inverse Tone MSE loss换回MSE loss。

I would like to ask why the inverse tone MSE loss causes the loss to become NaN? The the inverse tone function doesn't seem to cause this problem, but it does indeed fix NaN. CleanShot 2024-06-06 at 20 04 57@2x

cuiziteng commented 4 months ago

Thanks for your good question, in fact I think this maybe caused by arcsin function, I should add a clip process here.

I correct it now.

image
lucky9-cyou commented 4 months ago

Thanks for your good question, in fact I think this maybe caused by arcsin function, I should add a clip process here.

I correct it now.

image

Got it. Thanks for your reply.