Input image is large

[pribadihcr]

Hi,

What if the input image is large e.g 8192*14262, can the algorithm handle maybe get sub images but still larger than the template?. Thanks

[davidbuniat]

This method was used to compute correspondences between large images (50,000x50,000). To train the network fast enough, you would need to fit images into a GPU. We used ~200x200px for template and ~500x500px for source images. The maximum image width I reached was 1024px if I recall correctly.

To optimize the computation.

• Find an approximate location of the template in the large image
• Use subsampling of your images
• Construct a grid of matches

[pribadihcr]

@davidbuniat,

so the original image of your source image is 50Kx50K, and you did downsample to ~500x500 image size for training right?

[ClayZhang0403]

Hi，excuse me，How do I run your code program in my pc? Can I run your program through Cloudvolume?