Grid Sampling

To process equirectangular images quickly, equilib relies on grid sampling techniques. The goal of the project is to minimize external dependencies and take advantage of cuda and batch processing with torch for fast processing.

The library implements a variety of methods in numpy and torch:

• For torch, the built-in torch.nn.functional.grid_sample is fast and reliable. A pure-torch implementation is also provided and is highly customizable (though not necessarily as fast as the native function).
• For numpy, the implementations are faster than scipy and more robust than cv2.remap, and are just as customizable as the torch version.

It is also possible to pass scipy.ndimage.map_coordinates or cv2.remap as the sampling function via the override_func argument of grid_sample.

Notes

• By default, the numpy backend uses the pure-numpy implementation. Override it with scipy or cv2 via override_func.
• By default, the torch backend uses the official grid_sample.
• Benchmarking scripts live under benchmarks/. For example, benchmarks/equi2pers/numpy_run_baselines.py benchmarks the numpy equi2pers path against scipy and cv2.