Make YoloNASRDFLHead inherit a base YoloNASDFLHead

src/super_gradients/module_interfaces/exportable_obb_detector.py

@@ -51,7 +51,9 @@ def forward(self, predictions: Any) -> Tuple[Tensor, Tensor]:
 
         :return: Implementation of this method must return a tuple of two tensors (boxes, scores) with
         the following semantics:
-        - boxes - [B, N, 4]
+        - boxes - [B, N, 5] in (CX, CY, W, H, R) format.
+          R represents the angle oriented box in radians, counter-clockwise and following
+          OpenCV convention of angle of rotation for cv2.minAreaRect/cv2.boxPoints.
         - scores - [B, N, C]
         Where N is the maximum number of predictions per image (see self.get_num_pre_nms_predictions()),
         and C is the number of classes.

src/super_gradients/training/datasets/data_formats/obb/cxcywhr.py

@@ -24,6 +24,13 @@ def cxcywhr_to_poly(boxes: np.ndarray) -> np.ndarray:
 
 
 def poly_to_cxcywhr(poly: np.ndarray) -> np.ndarray:
+    """
+    Convert a polygon format to oriented bounding boxes in CX-CY-W-H-R format
+    :param poly: Input polygons in [N,...,4, 2] format
+    :return: [N,..., 5] Oriented bounding boxes in CX-CY-W-H-R format
+             R represents the angle oriented box in radians, counter-clockwise and following
+             OpenCV convention of angle of rotation for cv2.minAreaRect/cv2.boxPoints.
+    """
     shape = poly.shape
     if shape[-2:] != (4, 2):
         raise ValueError(f"Expected last two dimensions to be (4, 2), got {shape[-2:]}")

src/super_gradients/training/models/detection_models/yolo_nas_r/yolo_nas_r_ndfl_heads.py

@@ -37,6 +37,8 @@ class YoloNASRLogits:
     :param size_reduced: Tensor of shape [B, Anchors, 2] with predicted size distribution.
            None-multiplied by stride.
     :param angles: Tensor of shape [B, Anchors, 1] with predicted angles (in radians).
+           Angle is in range [-3*pi/4, pi/4]. Rotation is counter-clockwise and follows
+           OpenCV convention of angle of rotation for cv2.minAreaRect/cv2.boxPoints.
     :param offsets: Tensor of shape [B, Anchors, 2] with predicted offsets.
            Non-multiplied by stride.
     :param anchor_points: Tensor of shape [Anchors, 2] with anchor points.

src/super_gradients/training/models/detection_models/yolo_nas_r/yolo_nas_r_variants.py

@@ -89,6 +89,9 @@ def forward(self, inputs: Union[Tuple[Tensor, Tensor], YoloNASRLogits]):
 class YoloNASR(ExportableOBBDetectionModel, CustomizableDetector, SupportsInputShapeCheck):
     """
     YoloNAS-R model for Oriented Bounding Box (OBB) Detection.
+    The model is based on the YOLO-NAS architecture with minor modifications to support oriented bounding boxes.
+
+    Model supports export to ONNX and TensorRT via model.export(...) syntax.
     """
 
     def __init__(