Resolution r=1 make the scene blur.

[aruiplex]

Hi! I trained the bicycle scene with different parameters. When I set r=1, the scene blur than the default setting.

Here is the comparison figure:

right image set r=1, left image keeps default setting.

[Ph03n1xdust]

When you set the r=1, you force the training to use the very large resolution images. The bike images I had from somewhere are around 3x the resolution of the 1.6k width, which the authors aimed for. Consequently these images very likely contain much finer details than what the "default" settings or default starting point cloud was intended for.

I would try to set the lr-s lower, and train for more iterations. Also possibly allow the model to spawn more gaussians by setting the threshold lower and/or starting from a denser point cloud if you can obtain one.

But please take whatever I said with a grain of salt, this is just based on my very limited understanding of the model... And if you manage to test this way and see any improvement or worse results please report back so others also can get some info from here!

[HungNgoCT]

When you set the r=1, you force the training to use the very large resolution images. The bike images I had from somewhere are around 3x the resolution of the 1.6k width, which the authors aimed for. Consequently these images very likely contain much finer details than what the "default" settings or default starting point cloud was intended for.

I would try to set the lr-s lower, and train for more iterations. Also possibly allow the model to spawn more gaussians by setting the threshold lower and/or starting from a denser point cloud if you can obtain one.

But please take whatever I said with a grain of salt, this is just based on my very limited understanding of the model... And if you manage to test this way and see any improvement or worse results please report back so others also can get some info from here!

Hi @Ph03n1xdust
As you said: " Consequently these images very likely contain much finer details than what the "default" settings". I think if the input images contain finer details, we should expect the output should be finer. Why is it blur as @aruiplex mentioned?

[Ph03n1xdust]

" I think if the input images contain finer details, we should expect the output should be finer." - While in an ideal world you would be right, in practice these kind of models train somewhat weird. There are models where there is simply a limit of how good you can get, and throwing more data into it wont make it any better. In GS this limit is somewhat tied to the number of points.

Now the question is, did he start from the same number of points or way more? Did he tune the lr of the different components or not? Also while the currently implemented method takes care of creating "new gaussians", this is also very "handwavy" --> new gaussians are spawned when the accumulated positional gradient is high, but this is not necessarily the "best" way to look at it.

And because of all this you cannot really expect better and better results just by throwing more pixels on your model. Sure, if you start with a correct point cloud, and you tune your parameters correctly, then you should get better results.

See the FAQ about "extensive scenes", they also mentioned that for specific scenes the parameters should be tuned and in that case they also experienced some blur with no "reason".

So without actually seeing what he did exactly I can't really explain why is there a blur (and even if I knew exactly what he did, probably I would have no clue). But it's not fair to expect that by simply throwing more data on the same model you should get better results.

[HungNgoCT]

" I think if the input images contain finer details, we should expect the output should be finer." - While in an ideal world you would be right, in practice these kind of models train somewhat weird. There are models where there is simply a limit of how good you can get, and throwing more data into it wont make it any better. In GS this limit is somewhat tied to the number of points.

Now the question is, did he start from the same number of points or way more? Did he tune the lr of the different components or not? Also while the currently implemented method takes care of creating "new gaussians", this is also very "handwavy" --> new gaussians are spawned when the accumulated positional gradient is high, but this is not necessarily the "best" way to look at it.

And because of all this you cannot really expect better and better results just by throwing more pixels on your model. Sure, if you start with a correct point cloud, and you tune your parameters correctly, then you should get better results.

See the FAQ about "extensive scenes", they also mentioned that for specific scenes the parameters should be tuned and in that case they also experienced some blur with no "reason".

So without actually seeing what he did exactly I can't really explain why is there a blur (and even if I knew exactly what he did, probably I would have no clue). But it's not fair to expect that by simply throwing more data on the same model you should get better results.

Thank you for your explanation, @Ph03n1xdust

I can get your idea, although we still hope it should be better. Hope that @aruiplex will find a tune parameter set and share here.