# implementation inspired by https://github.com/chicleee/EDM/blob/main/demo_single_pair.ipynb
import torchvision.transforms as tfm
from safetensors.torch import load_file
from huggingface_hub import snapshot_download
from vismatch import THIRD_PARTY_DIR, BaseMatcher
from vismatch.utils import resize_to_divisible, add_to_path
add_to_path(THIRD_PARTY_DIR.joinpath("EDM"))
from src.edm import EDM
from src.config.default import get_cfg_defaults
from src.utils.misc import lower_config
[docs]
class EDMMatcher(BaseMatcher):
divisible_size = 32
def __init__(self, device="cpu", thresh=0.2, **kwargs):
super().__init__(device, **kwargs)
self.thresh = thresh
self.matcher = self.build_matcher()
[docs]
def build_matcher(self):
# Get default configurations
config = get_cfg_defaults()
config.merge_from_file(THIRD_PARTY_DIR / "EDM/configs/edm/outdoor/edm_base.py")
config.merge_from_file(THIRD_PARTY_DIR / "EDM/configs/data/megadepth_test_1500.py")
config.EDM.COARSE.MCONF_THR = self.thresh
config.EDM.COARSE.BORDER_RM = 2
config = lower_config(config)
matcher = EDM(config=config["edm"])
# Load model from HuggingFace
weights_path = f"{snapshot_download('vismatch/edm')}/edm.safetensors"
matcher.load_state_dict(load_file(weights_path))
return matcher.eval().to(self.device)
[docs]
def preprocess(self, img):
_, h, w = img.shape
orig_shape = h, w
img = resize_to_divisible(img, self.divisible_size)
return tfm.Grayscale()(img).unsqueeze(0), orig_shape
def _forward(self, img0, img1):
img0, img0_orig_shape = self.preprocess(img0)
img1, img1_orig_shape = self.preprocess(img1)
batch = {"image0": img0, "image1": img1}
self.matcher(batch)
mkpts0 = batch["mkpts0_f"]
mkpts1 = batch["mkpts1_f"]
H0, W0, H1, W1 = *img0.shape[-2:], *img1.shape[-2:]
mkpts0 = self.rescale_coords(mkpts0, *img0_orig_shape, H0, W0)
mkpts1 = self.rescale_coords(mkpts1, *img1_orig_shape, H1, W1)
return mkpts0, mkpts1, None, None, None, None