Fix KP Detection performance (#4270)

* fix kp det performance, change export, resize and augmentations

* move pad out of the resize

* remove second model, add decode_scores

* remove second config for RTMPose:

* rename kp template

* fix visual prompting

* fix torchvision unit test

* add unit test to rescale_kp

* delete redundant test

Author: kprokofi

src/otx/algo/keypoint_detection/rtmpose.py

@@ -43,13 +43,12 @@ def _exporter(self) -> OTXModelExporter:
         if self.explain_mode:
             msg = "Export with explain is not supported for RTMPose model."
             logger.warning(msg)
-
         return OTXNativeModelExporter(
             task_level_export_parameters=self._export_parameters,
             input_size=(1, 3, *self.input_size),
             mean=self.mean,
             std=self.std,
-            resize_mode="standard",
+            resize_mode="fit_to_window",
             pad_value=0,
             swap_rgb=False,
             via_onnx=True,
@@ -115,6 +114,7 @@ def _build_model(self, num_classes: int) -> RTMPose:
                 "sigma": sigma,
                 "normalize": False,
                 "use_dark": False,
+                "decode_scores": True,
             },
             gau_cfg={
                 "num_token": num_classes,

src/otx/algo/keypoint_detection/utils/simcc_label.py

@@ -67,7 +67,7 @@ def __init__(
         label_smooth_weight: float = 0.0,
         normalize: bool = True,
         use_dark: bool = False,
-        decode_visibility: bool = False,
+        decode_scores: bool = True,
         decode_beta: float = 150.0,
     ) -> None:
         self.input_size = input_size
@@ -76,7 +76,7 @@ def __init__(
         self.label_smooth_weight = label_smooth_weight
         self.normalize = normalize
         self.use_dark = use_dark
-        self.decode_visibility = decode_visibility
+        self.decode_scores = decode_scores
         self.decode_beta = decode_beta
 
         if isinstance(sigma, (float, int)):
@@ -170,13 +170,13 @@ def decode(self, simcc_x: np.ndarray, simcc_y: np.ndarray) -> tuple[np.ndarray,
 
         keypoints /= self.simcc_split_ratio
 
-        if self.decode_visibility:
+        if self.decode_scores:
             _, visibility = get_simcc_maximum(
                 simcc_x * self.decode_beta * self.sigma[0],
                 simcc_y * self.decode_beta * self.sigma[1],
                 apply_softmax=True,
             )
-            return keypoints, (scores, visibility)
+            return keypoints, visibility
         return keypoints, scores
 
     def _map_coordinates(

src/otx/core/data/dataset/keypoint_detection.py

@@ -108,7 +108,12 @@ def _get_item_impl(self, index: int) -> KeypointDetDataEntity | None:
             if len(keypoint_anns) > 0
             else np.zeros((0, len(self.label_info.label_names) * 2), dtype=np.float32)
         ).reshape(-1, 2)
-        keypoints_visible = np.minimum(1, keypoints)[..., 0]
+
+        keypoints_visible = (
+            (np.array([ann.visibility for ann in keypoint_anns]) > 1).reshape(-1).astype(np.int8)
+            if len(keypoint_anns) > 0 and hasattr(keypoint_anns[0], "visibility")
+            else np.minimum(1, keypoints)[..., 0]
+        )
 
         bbox_center = np.array(img_shape) / 2.0
         bbox_scale = np.array(img_shape)