SuperPoint and LightGlue by SimonovDmitry · Pull Request #7 · itlab-vision/feature-matching-mini

SimonovDmitry · 2026-03-29T21:01:55Z

No description provided.

ismukhin · 2026-03-31T19:04:01Z

src/super_point.py

+            raw_kp = processed['keypoints'].cpu().numpy()
+            raw_scores = processed['scores'].cpu().numpy()
+            raw_des = processed['descriptors'].cpu().numpy().astype(np.float32)


Вот здесь тонкий момент: возможно в какой-то момент захочется запускать инференс этих моделей на GPU, поэтому нужно будет сделать внешний параметр cpu или gpu, чтобы корректно управлять тензорами

ismukhin · 2026-03-31T19:10:12Z

src/super_point.py

+    def detect(self, img):
+        return self._forward(img)[0]
+
+    def compute(self, img, kp):
+        model_kp, model_des = self._forward(img)


Вроде обговаривали, что мы не инферим модель 2 раза. Мы проход по сетке делаем один раз и сохраняем результат в переменную, потом просто этот результат возвращаем (detect инферит, а compute возвращает ранее посчитанные дескрипторы). Только нужно будет подумать как создавать один и тот же экземпляр класса в 2 разных переменных

…f separate kp/des

valentina-kustikova · 2026-04-06T09:15:54Z

tests/test_super_point.py

+@pytest.fixture
+def load_img():
+    def _load(name, color=True):
+        path = Path(__file__).parent.parent / "test_data" / name


Здесь надо использовать os.path.join(...) для склеивания путей.

valentina-kustikova · 2026-04-06T09:32:35Z

src/matchers.py


    @abstractmethod
-    def _simple_match(self, des1, des2):
+    def match(self, **kwargs):


Вместо **kwargs передаем словарь.

valentina-kustikova · 2026-04-06T09:33:46Z

src/matchers.py

+        super().__init__(logger, matcher_name, descriptor_name)
+        self.mode = mode
+
+    def match(self, **kwargs):


Здесь тоже словарь типа {'des1': ..., 'des2': ...}

valentina-kustikova · 2026-04-06T09:35:00Z

src/matchers.py

-        bf = self._init_matcher()
-        return bf.match(des1, des2)
+class BFMatcher(OpenCVMatcher):
+    def __init__(self, logger, matcher_name, descriptor_name, **kwargs):


Вместо **kwargs используем словарь

valentina-kustikova · 2026-04-06T09:35:07Z

src/matchers.py

-    def __init__(self, logger, descriptor_method, mode='simple', **kwargs):
-        super().__init__(logger, descriptor_method, mode, **kwargs)
+class FLANNMatcher(OpenCVMatcher):
+    def __init__(self, logger, matcher_name, descriptor_name, **kwargs):


И здесь

valentina-kustikova · 2026-04-06T10:01:19Z

src/light_glue.py

+            LightGlueMatcher._shared_matchers[matcher_key] = (LightGlue(features=self.extractor_type)
+                                                              .eval().to(self._device))


Создать только один объект LightGlue

valentina-kustikova · 2026-04-06T10:02:30Z

src/light_glue.py

+        self._device = torch.device(device)
+        self.extractor_type = extractor_type.lower()
+
+        matcher_key = (self.extractor_type, self._device.type)


Удалить

valentina-kustikova · 2026-04-06T10:04:15Z

src/light_glue.py

+            matches01 = self._matcher(input_dict)
+            matches01 = rbd(matches01)
+
+            return matches01["matches"]


Нужно вернуть достоверности для соответствий

valentina-kustikova · 2026-04-06T10:08:09Z

src/super_point.py

+            kp = [cv.KeyPoint(x=float(p[0]), y=float(p[1]), size=8, response=float(s))
+                        for p, s in zip(raw_kp[mask], raw_scores[mask])]
+            des = raw_des[mask]


Вынести в отдельную функцию/класс PostProcessor в скрипт auxiliary.py перевод точек в формат cv.KeyPoint

valentina-kustikova · 2026-04-06T10:13:06Z

src/super_point.py

+
+    def compute(self, img, features):
+        kp, des = self._forward(img)
+        return {'kp': kp, 'des': des}


Аналогично ввести флаг состояния

…nd update all tests

SimonovDmitry · 2026-04-12T21:15:38Z

@ismukhin Исправили код, все тесты проходят, позапускали все алгоритмы, все работает.

ismukhin · 2026-04-13T09:41:56Z

samples/match_cv.py



+def build_detector_config(args):
+    config = {}


ismukhin · 2026-04-13T09:42:12Z

samples/match_cv.py

+
+
+def build_descriptor_config(args):
+    config = {}


ismukhin · 2026-04-13T09:42:19Z

samples/match_cv.py

+
+
+def build_matcher_config(args):
+    config = {}


ismukhin · 2026-04-13T09:42:29Z

samples/match_cv.py

+
+
+def build_preprocessor_config(args):
+    config = {}


ismukhin · 2026-04-13T09:45:24Z

samples/match_cv.py

+def build_preprocessor_config(args):
+    config = {}
+    if args.pre_device is not None:
+        config['device'] = args.pre_device


Непонятно зачем здесь device, нужно гарантировать, что все операции выполняются на одном устройстве

ismukhin · 2026-04-13T10:11:57Z

src/lightglue_pipeline.py

@@ -0,0 +1,85 @@
+import cv2 as cv


lightglue_pipeline.py

ismukhin · 2026-04-13T10:13:34Z

src/preprocessor.py

+        if algo in self._VALID_FORMATS:
+            return algo
+
+        if algo in NEURAL_ALGORITHMS:


DNN_{всё остальное}

ismukhin · 2026-04-13T10:14:05Z

src/preprocessor.py

+from src.converter import Converter
+
+
+class PreProcessor:


Preprocessor

ismukhin · 2026-04-13T10:18:39Z

src/super_point.py

+        else:
+            self._device = torch.device(device)
+
+        if SuperPoint._shared_model is None:


SuperPoint._model

ismukhin · 2026-04-13T10:18:48Z

src/super_point.py

+            self._device = torch.device(device)
+
+        if SuperPoint._shared_model is None:
+            SuperPoint._shared_processor = AutoImageProcessor.from_pretrained(


SuperPoint._image_processor

…results

add SuperPoint, LightGlue implementation and tests for SuperPoint

de51df3

ismukhin reviewed Mar 31, 2026

View reviewed changes

refactored Matcher,SuperPoint and LightGlue to handle dicts instead o…

723ef6e

…f separate kp/des

valentina-kustikova reviewed Apr 6, 2026

View reviewed changes

SimonovDmitry added 3 commits April 12, 2026 22:46

add config, preprocessor, data converters, fix SuperPoint/LightGlue a…

b8f2ca5

…nd update all tests

removed empty lines for CI

d830e47

removed the last indent lines

f37edee

ismukhin reviewed Apr 13, 2026

View reviewed changes

SimonovDmitry and others added 3 commits April 14, 2026 21:10

refactor: add MODE dict, fix naming, share device param, always show …

3b4a646

…results

code style: corrected variable names

4c304e0

temp fix for ci

3ba6d2f

ismukhin approved these changes Apr 14, 2026

View reviewed changes

ismukhin merged commit b2f3e36 into itlab-vision:main Apr 14, 2026
3 checks passed

		LightGlueMatcher._shared_matchers[matcher_key] = (LightGlue(features=self.extractor_type)
		.eval().to(self._device))

Conversation

SimonovDmitry commented Mar 29, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SimonovDmitry commented Apr 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants