update: gim

common/utils.py

@@ -14,7 +14,7 @@ from .viz import draw_matches, fig2im, plot_images, plot_color_line_matches
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
 DEFAULT_SETTING_THRESHOLD = 0.1
-DEFAULT_SETTING_MAX_FEATURES = 2000
+DEFAULT_SETTING_MAX_FEATURES = 4096
 DEFAULT_DEFAULT_KEYPOINT_THRESHOLD = 0.01
 DEFAULT_ENABLE_RANSAC = True
 DEFAULT_RANSAC_METHOD = "USAC_MAGSAC"
@@ -44,12 +44,7 @@ def gen_examples():
         "gim",
         "gim",
         "gim",
-        "disk+lightglue",
-        "loftr",
-        "disk",
-        "d2net",
-        "superpoint+superglue",
-        "disk+dualsoftmax",
+        "gim",
     ]
 
     def gen_images_pairs(path: str, count: int = 5):
@@ -67,8 +62,9 @@ def gen_examples():
     pairs = gen_images_pairs(path, len(example_matchers))
     gim_pairs = [('datasets/gim/0a.png', 'datasets/gim/0b.png'),
                  ('datasets/gim/1a.png', 'datasets/gim/1b.png'),
-                 ('datasets/gim/2a.png', 'datasets/gim/2b.png')]
-    pairs = gim_pairs + pairs
+                 ('datasets/gim/2a.png', 'datasets/gim/2b.png'),
+                 ('datasets/gim/3a.png', 'datasets/gim/3b.png')]
+    pairs = gim_pairs
     match_setting_threshold = DEFAULT_SETTING_THRESHOLD
     match_setting_max_features = DEFAULT_SETTING_MAX_FEATURES
     detect_keypoints_threshold = DEFAULT_DEFAULT_KEYPOINT_THRESHOLD
@@ -77,7 +73,7 @@ def gen_examples():
     ransac_confidence = DEFAULT_RANSAC_CONFIDENCE
     ransac_max_iter = DEFAULT_RANSAC_MAX_ITER
     input_lists = []
-    for pair, mt in zip(pairs, example_matchers):
+    for pair, mt in zip(gim_pairs, example_matchers):
         input_lists.append(
             [
                 pair[0],

hloc/match_dense.py

@@ -14,16 +14,11 @@ confs = {
         "model": {
             "name": "gim",
             "weights": "gim_dkm_100h.ckpt",
-            "max_keypoints": 2000,
+            "max_keypoints": 4096,
             "match_threshold": 0.2,
         },
         "preprocessing": {
             "grayscale": False,
-            "force_resize": True,
-            "resize_max": 1024,
-            "width": 80,
-            "height": 60,
-            "dfactor": 8,
         },
     },
     "loftr": {
@@ -279,12 +274,10 @@ def match(model, path_0, path_1, conf):
 def match_images(model, image_0, image_1, conf, device="cpu"):
     default_conf = {
         "grayscale": True,
-        "resize_max": 1024,
+        "resize_max": 8192,
         "dfactor": 8,
         "cache_images": False,
         "force_resize": False,
-        "width": 320,
-        "height": 240,
     }
 
     def preprocess(image: np.ndarray):

hloc/matchers/gim.py

@@ -72,21 +72,38 @@ class GIM(BaseModel):
         # )
 
         image0, image1 = data['image0'], data['image1']
-        orig_width = image0.shape[3]
-        orig_height = image0.shape[2]
+        print('1.', 'image0', image0.shape, 'image1', image1.shape)
+        orig_width0 = image0.shape[3]
+        orig_height0 = image0.shape[2]
+        orig_width1 = image1.shape[3]
+        orig_height1 = image1.shape[2]
         aspect_ratio = 896 / 672
-        new_width = max(orig_width, int(orig_height * aspect_ratio))
-        new_height = max(orig_height, int(orig_width / aspect_ratio))
-        pad_height = new_height - orig_height
-        pad_width = new_width - orig_width
-        pad_top = pad_height // 2
-        pad_bottom = pad_height - pad_top
-        pad_left = pad_width // 2
-        pad_right = pad_width - pad_left
-        image0 = torch.nn.functional.pad(image0, (pad_left, pad_right, pad_top, pad_bottom))
-        image1 = torch.nn.functional.pad(image1, (pad_left, pad_right, pad_top, pad_bottom))
+
+        new_width0 = max(orig_width0, int(orig_height0 * aspect_ratio))
+        new_height0 = max(orig_height0, int(orig_width0 / aspect_ratio))
+        new_width1 = max(orig_width1, int(orig_height1 * aspect_ratio))
+        new_height1 = max(orig_height1, int(orig_width1 / aspect_ratio))
+
+        new_width = max(new_width0, new_width1)
+        new_height = max(new_height0, new_height1)
+
+        pad_height0 = new_height - orig_height0
+        pad_width0 = new_width - orig_width0
+        pad_height1 = new_height - orig_height1
+        pad_width1 = new_width - orig_width1
+        pad_top0 = pad_height0 // 2
+        pad_bottom0 = pad_height0 - pad_top0
+        pad_left0 = pad_width0 // 2
+        pad_right0 = pad_width0 - pad_left0
+        pad_top1 = pad_height1 // 2
+        pad_bottom1 = pad_height1 - pad_top1
+        pad_left1 = pad_width1 // 2
+        pad_right1 = pad_width1 - pad_left1
+        image0 = torch.nn.functional.pad(image0, (pad_left0, pad_right0, pad_top0, pad_bottom0))
+        image1 = torch.nn.functional.pad(image1, (pad_left1, pad_right1, pad_top1, pad_bottom1))
+        print('2.', 'image0', image0.shape, 'image1', image1.shape)
         dense_matches, dense_certainty = self.net.match(image0, image1)
-        sparse_matches, mconf = self.net.sample(dense_matches, dense_certainty, 2048)
+        sparse_matches, mconf = self.net.sample(dense_matches, dense_certainty, self.conf["max_keypoints"])
         height0, width0 = image0.shape[-2:]
         height1, width1 = image1.shape[-2:]
         kpts0 = sparse_matches[:, :2]
@@ -95,17 +112,17 @@ class GIM(BaseModel):
         kpts1 = torch.stack((width1 * (kpts1[:, 0] + 1) / 2, height1 * (kpts1[:, 1] + 1) / 2), dim=-1, )
         b_ids, i_ids = torch.where(mconf[None])
         # before padding
-        kpts0 -= kpts0.new_tensor((pad_left, pad_top))[None]
-        kpts1 -= kpts1.new_tensor((pad_left, pad_top))[None]
+        kpts0 -= kpts0.new_tensor((pad_left0, pad_top0))[None]
+        kpts1 -= kpts1.new_tensor((pad_left1, pad_top1))[None]
         mask = (kpts0[:, 0] > 0) & \
                (kpts0[:, 1] > 0) & \
                (kpts1[:, 0] > 0) & \
                (kpts1[:, 1] > 0)
         mask = mask & \
-               (kpts0[:, 0] <= (orig_width - 1)) & \
-               (kpts1[:, 0] <= (orig_width - 1)) & \
-               (kpts0[:, 1] <= (orig_height - 1)) & \
-               (kpts1[:, 1] <= (orig_height - 1))
+               (kpts0[:, 0] <= (orig_width0 - 1)) & \
+               (kpts1[:, 0] <= (orig_width1 - 1)) & \
+               (kpts0[:, 1] <= (orig_height0 - 1)) & \
+               (kpts1[:, 1] <= (orig_height1 - 1))
         pred = {
             'keypoints0': kpts0[i_ids],
             'keypoints1': kpts1[i_ids],