Bad result of nerfacto on my dataset

[asedfrgt]

Thank for your grateful work. I use the default nerf acto model on my data. The trainning psnr is 27.8db, but the testing psnr is only 20db. Why is it overfting?

[tancik]

Can you share an example render results? Bad results can be caused by a number of factors such as not enough captures of the scene, motion blur, dynamic scene elements, changes in exposure/appearance. It is also worth noting that psnr is not a great metric for real scene. The appearance embeddings and pose optimization will improve the training psnr and hurt the test psnr, however the perceived quality if you render out a video will look better as it will have fewer floating artifacts and sharper results.

[asedfrgt]

Thanks for your answer. I don't use the pose optimatimzition and appearance embending. The test ing psnr is also 20 psnr, training psnr 26db.

[asedfrgt]

When I use the instant-ngp method, the train and test psnr are close to being equal.

[tancik]

Can you share some example data, it may help with debugging.

[asedfrgt]

Hi, I test the nerfactor sf dataset. The train-num-rays-per-batch is 8192. Orientation-method is pca. After 33000 iterations, the training psnr is 32db, but testing psnr is about 20psnr.

[asedfrgt]

The new view image has no details

[tancik]

Oh, I'm only now realizing that the sf dataset is one what we provided. This dataset is bad and has been removed in the latest release.