Merge pull request rpautrat#11 from rpautrat/improve_homographies

Improve homographies

superpoint/configs/magic-point_coco_export.yaml

@@ -6,9 +6,9 @@ model:
     name: 'magic_point'
     batch_size: 50
     learning_rate: 0.001
-    detection_threshold: 0.015 # 1/65
+    detection_threshold: 0.001 # 0.015
     nms: 4
-    top_k: 300
+    top_k: 1000
     homography_adaptation:
         num: 100
         aggregation: 'sum'
@@ -19,8 +19,8 @@ model:
             scaling: true
             perspective: true
             scaling_amplitude: 0.1
-            perspective_amplitude_x: 0.15
-            perspective_amplitude_y: 0.15
+            perspective_amplitude_x: 0.1
+            perspective_amplitude_y: 0.1
             allow_artifacts: true
             patch_ratio: 0.5
 eval_iter: -1

superpoint/configs/magic-point_coco_train.yaml

@@ -1,6 +1,6 @@
 data:
     name: 'coco'
-    labels: null # Complete with your export labels
+    labels: outputs/mp_synth-v7_ha2-100_no-topk # Complete with your export labels
     cache_in_memory: false
     validation_size: 192
     augmentation:
@@ -26,19 +26,19 @@ data:
                 scaling: true
                 perspective: true
                 scaling_amplitude: 0.2
-                perspective_amplitude_x: 0.4
-                perspective_amplitude_y: 0.3
+                perspective_amplitude_x: 0.1
+                perspective_amplitude_y: 0.1
                 patch_ratio: 0.85
-                max_angle: 3.14
+                max_angle: 1.57
                 allow_artifacts: true
             valid_border_margin: 3
 model:
     name: 'magic_point'
     batch_size: 32
     eval_batch_size: 32
     learning_rate: 0.001
-    detection_threshold: 0.015 # 1/65
+    detection_threshold: 0.001 # 0.015
     nms: 4
-    top_k: 300
+    # top_k: 300
 train_iter: 200000
 validation_interval: 1000

superpoint/configs/magic-point_repeatability.yaml

@@ -3,17 +3,17 @@ data:
     dataset: 'hpatches'  # 'coco' 'hpatches'
     alteration: 'v'  # 'all' 'i' 'v'
     preprocessing:
-        resize: [240, 320]  # False for coco
+        resize: [240, 320]  # [240, 320] for HPatches and False for coco
 model:
     name: 'magic_point'
     batch_size: 50
     eval_batch_size: 50
     learning_rate: 0.001
     detection_threshold: 0.001
     nms: 4
-    top_k: 300
+    top_k: 1000
     homography_adaptation:
-        num: 100
+        num: 0
         aggregation: 'sum'
         filter_counts: 0
         homographies:
@@ -22,7 +22,8 @@ model:
             scaling: true
             perspective: true
             scaling_amplitude: 0.1
-            perspective_amplitude_x: 0.05
-            perspective_amplitude_y: 0.05
-            allow_artifacts: false
+            perspective_amplitude_x: 0.1
+            perspective_amplitude_y: 0.1
+            allow_artifacts: true
+            patch_ratio: 0.5
 eval_iter: 1000

superpoint/evaluations/detector_evaluation.py

@@ -157,7 +157,7 @@ def filter_keypoints(points, shape):
 
     def keep_true_keypoints(points, H, shape):
         """ Keep only the points whose warped coordinates by H
-        are still inside shape.. """
+        are still inside shape. """
         warped_points = warp_keypoints(points[:, [1, 0]], H)
         warped_points[:, [0, 1]] = warped_points[:, [1, 0]]
         mask = (warped_points[:, 0] >= 0) & (warped_points[:, 0] < shape[0]) &\
@@ -189,7 +189,6 @@ def select_k_best(points, k):
         warped_keypoints = np.stack([warped_keypoints[0],
                                      warped_keypoints[1],
                                      warped_prob], axis=-1)
-        warped_keypoints = select_k_best(warped_keypoints, keep_k_points)  # TODO
         warped_keypoints = keep_true_keypoints(warped_keypoints, np.linalg.inv(H),
                                                data['prob'].shape)
 
@@ -198,12 +197,11 @@ def select_k_best(points, k):
         true_warped_keypoints = np.stack([true_warped_keypoints[:, 1],
                                           true_warped_keypoints[:, 0],
                                           prob], axis=-1)
-        true_warped_keypoints = select_k_best(true_warped_keypoints, keep_k_points)  # TODO
         true_warped_keypoints = filter_keypoints(true_warped_keypoints, shape)
 
         # Keep only the keep_k_points best predictions
-        # warped_keypoints = select_k_best(warped_keypoints, keep_k_points)
-        # true_warped_keypoints = select_k_best(true_warped_keypoints, keep_k_points)
+        warped_keypoints = select_k_best(warped_keypoints, keep_k_points)
+        true_warped_keypoints = select_k_best(true_warped_keypoints, keep_k_points)
 
         # Compute the repeatability
         N1 = true_warped_keypoints.shape[0]

superpoint/models/homographies.py

@@ -15,7 +15,7 @@
             'scaling': True,
             'perspective': True,
             'scaling_amplitude': 0.1,
-            'perspective_amplitude_x': 0.2,
+            'perspective_amplitude_x': 0.1,
             'perspective_amplitude_y': 0.1,
             'patch_ratio': 0.5,
             'max_angle': pi,
@@ -104,8 +104,8 @@ def step(i, probs, counts, images):
 
 def sample_homography(
         shape, perspective=True, scaling=True, rotation=True, translation=True,
-        n_scales=5, n_angles=25, scaling_amplitude=0.1, perspective_amplitude_x=0.2,
-        perspective_amplitude_y=0.1, patch_ratio=0.5, max_angle=pi,
+        n_scales=5, n_angles=25, scaling_amplitude=0.1, perspective_amplitude_x=0.1,
+        perspective_amplitude_y=0.1, patch_ratio=0.5, max_angle=pi/2,
         allow_artifacts=False, translation_overflow=0.):
     """Sample a random valid homography.
 
@@ -122,8 +122,15 @@ def sample_homography(
         scaling: A boolean that enables the random scaling of the patch.
         rotation: A boolean that enables the random rotation of the patch.
         translation: A boolean that enables the random translation of the patch.
-        n_scales: the number of tentative scales that are sampled when scaling.
-        n_angles: the number of tentatives angles that are sampled when rotating.
+        n_scales: The number of tentative scales that are sampled when scaling.
+        n_angles: The number of tentatives angles that are sampled when rotating.
+        scaling_amplitude: Controls the amount of scale.
+        perspective_amplitude_x: Controls the perspective effect in x direction.
+        perspective_amplitude_y: Controls the perspective effect in y direction.
+        patch_ratio: Controls the size of the patches used to create the homography.
+        max_angle: Maximum angle used in rotations.
+        allow_artifacts: A boolean that enables artifacts when applying the homography.
+        translation_overflow: Amount of border artifacts caused by translation.
 
     Returns:
         A `Tensor` of shape `[1, 8]` corresponding to the flattened homography transform.
@@ -142,9 +149,14 @@ def sample_homography(
         if not allow_artifacts:
             perspective_amplitude_x = min(perspective_amplitude_x, margin)
             perspective_amplitude_y = min(perspective_amplitude_y, margin)
-        pts2 += tf.concat([tf.truncated_normal([4, 1], 0., perspective_amplitude_x/2),
-                           tf.truncated_normal([4, 1], 0., perspective_amplitude_y/2)],
-                          axis=1)
+        perspective_displacement = tf.truncated_normal([1], 0., perspective_amplitude_y)
+        h_displacement_left = tf.truncated_normal([1], 0., perspective_amplitude_x)
+        h_displacement_right = tf.truncated_normal([1], 0., perspective_amplitude_x)
+        pts2 += tf.stack([tf.concat([h_displacement_left, perspective_displacement], 0),
+                          tf.concat([h_displacement_left, -perspective_displacement], 0),
+                          tf.concat([h_displacement_right, perspective_displacement], 0),
+                          tf.concat([h_displacement_right, -perspective_displacement],
+                                    0)])
 
     # Random scaling
     # sample several scales, check collision with borders, randomly pick a valid one