Further efforts to reproduce loftr hpatches

configs/loftr.yml

@@ -3,12 +3,17 @@ default: &default
     ckpt: 'pretrained/loftr/outdoor_ds.ckpt'
     match_threshold: 0.2
     imsize: -1
-    no_matches_upscale: False
+    aspect_ratio: -1
+    no_match_upscale: False
+    nbr_matches: -1
 example:
     <<: *default
     match_threshold: 0.5
     imsize: -1
 hpatch:        
     <<: *default
-    imsize: 480
+    match_threshold: 0.5
+    imsize: 640
+    aspect_ratio: 1.333333  # images will be 640x480 or 480x640
     no_match_upscale: True
+    nbr_matches: 1000

immatch/modules/loftr.py

@@ -14,8 +14,10 @@ def __init__(self, args):
             args = Namespace(**args)
 
         self.imsize = args.imsize        
+        self.aspect_ratio = args.aspect_ratio
         self.match_threshold = args.match_threshold
         self.no_match_upscale = args.no_match_upscale
+        self.nbr_matches = args.nbr_matches
 
         # Load model
         conf = dict(default_cfg)
@@ -34,7 +36,7 @@ def __init__(self, args):
 
     def load_im(self, im_path):
         return load_gray_scale_tensor_cv(
-            im_path, self.device, imsize=self.imsize, dfactor=8
+            im_path, self.device, imsize=self.imsize, aspect_ratio=self.aspect_ratio, dfactor=8
         )
 
     def match_inputs_(self, gray1, gray2):
@@ -43,6 +45,12 @@ def match_inputs_(self, gray1, gray2):
         kpts1 = batch['mkpts0_f'].cpu().numpy()
         kpts2 = batch['mkpts1_f'].cpu().numpy()
         scores = batch['mconf'].cpu().numpy()
+        if self.nbr_matches and self.nbr_matches > 0:
+            ids = np.argsort(scores)[-self.nbr_matches:]
+            kpts1 = kpts1[ids]
+            kpts2 = kpts2[ids]
+            scores = scores[ids]
+
         matches = np.concatenate([kpts1, kpts2], axis=1)
         return matches, kpts1, kpts2, scores
 

immatch/utils/data_io.py

@@ -10,10 +10,19 @@ def lprint(ms, log=None):
         log.write(ms+'\n')
         log.flush()
 
-def resize_im(wo, ho, imsize=None, dfactor=1, value_to_scale=max):
+def resize_im(wo, ho, imsize=None, aspect_ratio=None, dfactor=1):
     wt, ht = wo, ho
-    if imsize and value_to_scale(wo, ho) > imsize and imsize > 0:
-        scale = imsize / value_to_scale(wo, ho)
+    if aspect_ratio and aspect_ratio > 0:
+        if not imsize or imsize < 0:
+            raise ValueError('imsize must be given when using fixed aspect_ratio')
+        max_wh = imsize
+        min_wh = round(imsize / aspect_ratio)
+        if wo < ho:
+            wt, ht = min_wh, max_wh
+        else:
+            wt, ht = max_wh, min_wh
+    elif imsize and max(wo, ho) > imsize and imsize > 0:
+        scale = imsize / max(wo, ho)
         ht, wt = int(round(ho * scale)), int(round(wo * scale))
 
     # Make sure new sizes are divisible by the given factor
@@ -41,11 +50,11 @@ def load_gray_scale_tensor(im_path, device, imsize=None, dfactor=1):
     gray = transforms.functional.to_tensor(gray).unsqueeze(0).to(device)
     return gray, scale
 
-def load_gray_scale_tensor_cv(im_path, device, imsize=None, dfactor=1):
+def load_gray_scale_tensor_cv(im_path, device, imsize=None, aspect_ratio=None, dfactor=1):
     # Used for LoFTR
     im = cv2.imread(im_path, cv2.IMREAD_GRAYSCALE)
     ho, wo = im.shape
-    wt, ht, scale = resize_im(wo, ho, imsize=imsize, dfactor=dfactor, value_to_scale=min)
+    wt, ht, scale = resize_im(wo, ho, imsize=imsize, aspect_ratio=aspect_ratio, dfactor=dfactor)
     im = cv2.resize(im, (wt, ht))
     im = transforms.functional.to_tensor(im).unsqueeze(0).to(device)
     return im, scale

immatch/utils/hpatches_helper.py

@@ -197,7 +197,7 @@ def eval_hpatches(
             if 'homography' in task:
                 try:
                     if 'cv' in h_solver:
-                        H_pred, inliers = cv2.findHomography(matches[:, :2], matches[:, 2:4], cv2.RANSAC, ransac_thres)
+                        H_pred, inliers = cv2.findHomography(matches[:, :2], matches[:, 2:4], cv2.RANSAC, ransac_thres, confidence=0.99999)
                     else:
                         H_pred, inliers = pydegensac.findHomography(matches[:, :2], matches[:, 2:4], ransac_thres)
                 except: