Dust3r Unable to Deal with Many Images, Flooding the RAM
eyildiz-ugoe opened this issue · comments
I'm using the provided demo script to feed in around 300 images (with 2K resolution, though they get reduced to 512x512 anyway) to get a precise 3D model of the scene. The inference takes forever, I've launched it like 1h ago and it says it is still at 30% on inference. Apparently there are 107256 image pairs (?) and it is unable to process all. I tried it a few times, it always comes to 30% and then it gets stuck, the screen is not interactable anymore. My computer is not even bad, I have RTX 3090 and a decent CPU. So it's not about the hardware per se. For the past 10 mins or so, there is no change in the progress bar, the terminal is frozen. My only option is to do a hardware reset, that is, to literally cut the power off. That's just straight up bad, isn't it?
I've executed htop on another terminal before I started the execution. What happens is that your code, once launched, gradually occupies massive amount of RAM. I got like 256GB of RAM and all of that is consumed by Dust3r after 30 mins, that can't be normal. Either the demo parameters are off, or the implementation has a terrible bottleneck the authors are not aware of.
In any case, this is a problem. Just letting you know so that you refactor it on your side.
Yes, it is a problem, and you cannot work with many images. They are working on an optimized solution for this issue. By the way, for scene_graph, you can use 'swin' or 'oneref' to find a smaller number of pairs, resulting in less computation.