OOM

[StarsTesla]

Hi, I tried this code on my pc, and I used 400 images, this seems OOM, does this code load all images to GPU and run 400 images in mean time? COLMAP seems not OOM on over 400 images

[StarsTesla]

I reduce to 200 images, and use 4090, still OOM

[dcharatan]

You can try appending +experiment=low_memory to the command you use to run FlowMap. However, FlowMap isn't designed to run on video sequences of many hundreds or thousands of frames. There are probably ways to address this (see #4 and #13 for more details), but we left that for future work.

[kk6398]

You can try appending +experiment=low_memory to the command you use to run FlowMap. However, FlowMap isn't designed to run on video sequences of many hundreds or thousands of frames. There are probably ways to address this (see #4 and #13 for more details), but we left that for future work.

The "image_shape: 10800" in ./config/experiment/low_memory.yaml indicates that the shape of image was cropped to "10800"? What is the specific height and width?

[dcharatan]

If the image_shape is just one number, it's the approximate total number of pixels (so width times height).

[kk6398]

If the image_shape is just one number, it's the approximate total number of pixels (so width times height).

Yep, for example, the image resolution of mipnerf360_garden scene is 51873316, which the total number of pixels is much larger than 10800. Is it cropped from 51873316 to 10800?

[dcharatan]

Yes, FlowMap will resize the input images until they're roughly at the target resolution. There are two resolutions that FlowMap uses:

1. The resolution at which the optimization runs, which is determined by the image_shape parameter. This is a relatively low resolution, since running a CNN on high-resolution images is very expensive.
2. The resolution at which optical flow and point tracks are calculated, which is generally 16x-25x as large (4x to 5x as large in each dimension, determined by flow_scale_multiplier). High-resolution images are needed for flow and tracks to be accurate.

Once FlowMap has resized the images as determined by image_shape, it will crop them until their width and height are divisible by 32. This is necessary for compatibility with RAFT.

[kk6398]

Yes, FlowMap will resize the input images until they're roughly at the target resolution. There are two resolutions that FlowMap uses:

1. The resolution at which the optimization runs, which is determined by the image_shape parameter. This is a relatively low resolution, since running a CNN on high-resolution images is very expensive.
2. The resolution at which optical flow and point tracks are calculated, which is generally 16x-25x as large (4x to 5x as large in each dimension, determined by flow_scale_multiplier). High-resolution images are needed for flow and tracks to be accurate.

Once FlowMap has resized the images as determined by image_shape, it will crop them until their width and height are divisible by 32. This is necessary for compatibility with RAFT.

Hi, ①I take a good look at the code. I discovery that the default image_shape is 43200(180*240) for demo in "overfit.yaml". I wonder whether the resolution is all 43200 at inference process in condition of without using "+experiment=low_memory".

②Besides, the image_shape for co3d/re10k at training stage is default [300,400]. What is the specific resolution of co3d/re10k? And how many memories cost of [300,400]? [300,400] is the experience value that we debug?
Looking forward your reply.