Prevent zero-sized saliency map in tiling if tile size is too big (#2452)

* Prevent zero-sized saliency map in tiling if tile size is too big

* Prevent zero-sized saliency in tiling (PyTorch)

* Add unit tests for Tiler merge features methods

---------

Co-authored-by: Galina <[email protected]>

Author: sovrasov

src/otx/algorithms/detection/adapters/mmdet/datasets/tiling.py

@@ -527,8 +527,8 @@ def merge_maps(self, saliency_maps: Union[List[List[np.ndarray]], List[np.ndarra
 
         for orig_image in self.cached_results:
             img_idx = orig_image["index"]
-            ratios[img_idx] = np.array([feat_h, feat_w]) / self.tile_size
             image_h, image_w = orig_image["height"], orig_image["width"]
+            ratios[img_idx] = np.array([feat_h / min(self.tile_size, image_h), feat_w / min(self.tile_size, image_w)])
 
             image_map_h = int(image_h * ratios[img_idx][0])
             image_map_w = int(image_w * ratios[img_idx][1])

src/otx/api/utils/tiler.py

@@ -325,14 +325,13 @@ def merge_maps(self, features: List) -> np.ndarray:
         Returns:
             merged_maps (np.ndarray): Merged saliency maps for entire image.
         """
-
         (_, image_saliency_map), image_meta = features[0]
 
         num_classes, feat_h, feat_w = image_saliency_map.shape
         dtype = image_saliency_map[0][0].dtype
 
         image_h, image_w, _ = image_meta["original_shape"]
-        ratio = np.array([feat_h, feat_w]) / self.tile_size
+        ratio = np.array([feat_h / min(self.tile_size, image_h), feat_w / min(self.tile_size, image_w)])
 
         image_map_h = int(image_h * ratio[0])
         image_map_w = int(image_w * ratio[1])

tests/unit/algorithms/detection/tiling/test_tiling_tile_classifier.py

@@ -78,8 +78,12 @@ def test_openvino_sync(self, mocker):
         mocked_model.return_value = mocker.MagicMock(spec=OTXMaskRCNNModel, model_adapter=adapter_mock)
         params = DetectionConfig(header=self.hyper_parameters.header)
         ov_mask_inferencer = OpenVINOMaskInferencer(params, self.label_schema, "")
+        original_shape = (self.dataset[0].media.width, self.dataset[0].media.height, 3)
         ov_mask_inferencer.model.resize_mask = False
-        ov_mask_inferencer.model.preprocess.return_value = ({"foo": "bar"}, {"baz": "qux"})
+        ov_mask_inferencer.model.preprocess.return_value = (
+            {"foo": "bar"},
+            {"baz": "qux", "original_shape": original_shape},
+        )
         ov_mask_inferencer.model.postprocess.return_value = (
             np.array([], dtype=np.float32),
             np.array([], dtype=np.uint32),
@@ -93,6 +97,10 @@ def test_openvino_sync(self, mocker):
         mock_predict = mocker.patch.object(
             ov_inferencer.tiler.classifier, "infer_sync", return_value={"tile_prob": 0.5}
         )
+        ov_inferencer.tiler.model.infer_sync.return_value = {
+            "feature_vector": np.zeros((1, 5), dtype=np.float32),
+            "saliency_map": np.zeros((1, 1, 2, 2), dtype=np.float32),
+        }
         mocker.patch.object(OpenVINODetectionTask, "load_inferencer", return_value=ov_inferencer)
         ov_task = OpenVINODetectionTask(self.task_env)
         ov_task.inferencer.predict = partial(ov_task.inferencer.predict, mode="sync")