Visualizer._convert_boxes produces N x 1 x 4 matrix instead of N x 4

[marie-r]

In the 0.6 detectron2 release, I am using the Visualizer to verify ground truth labels using Visualizer.draw_dataset_dict. My annotations have XYXY bounding boxes and draw_dataset_dict pulls out into a list of numpy arrays, all 1 x 4, e.g.:

[array([[553., 209., 773., 227.]], dtype=float32), array([[ 951., 250., 1341., 290.]], dtype=float32), array([[1974., 207., 2227., 227.]], dtype=float32), array([[2390., 251., 2771., 289.]], dtype=float32)]

When Visualizer.draw_dataset_dict calls Visualizer.overlay_instances, it passes in the list of boxes (line 603). One of the first operations in overlay_instances is to process the boxes by calling _convert_boxes. (line 1227)

def _convert_boxes(self, boxes): """ Convert different format of boxes to an NxB array, where B = 4 or 5 is the box dimension. """ if isinstance(boxes, Boxes) or isinstance(boxes, RotatedBoxes): return boxes.tensor.detach().numpy() else: return np.array(boxes)

The list falls into the else leg. With numpy 2.2.6 this produces a 4 x 1 x 4 tensor which causes problems down the road. While it can easily be fixed by changing np.array to np.vstack, I was wondering if my data were wrong (I'm well versed in ML, not in detectron2).

Thanks - Marie