From ae43903177552858e1ef41029cdd6e115cd71c68 Mon Sep 17 00:00:00 2001
From: dcarron <daniel.carron@idiap.ch>
Date: Fri, 19 Jan 2024 13:50:47 +0100
Subject: [PATCH] Moved computation of binary mask to its own function
---
.../engine/saliency/interpretability.py | 42 +++++++++++++++----
1 file changed, 35 insertions(+), 7 deletions(-)
diff --git a/src/ptbench/engine/saliency/interpretability.py b/src/ptbench/engine/saliency/interpretability.py
index d56aa846..1275ea5d 100644
--- a/src/ptbench/engine/saliency/interpretability.py
+++ b/src/ptbench/engine/saliency/interpretability.py
@@ -260,6 +260,40 @@ def _compute_proportional_energy(
return float(numpy.sum(saliency_map * gt_mask) / denominator) # type: ignore
+def _compute_binary_mask(
+ gt_bboxes: BoundingBoxes,
+ saliency_map: numpy.typing.NDArray[numpy.double],
+) -> numpy.typing.NDArray[numpy.bool_]:
+ """Computes a binary mask for the saliency map using BoundingBoxes.
+
+ The binary_mask will be ON/True where the gt boxes are located.
+
+ Parameters
+ ----------
+ gt_bboxes
+ Ground-truth bounding boxes in the format ``(x, y, width,
+ height)``.
+
+ saliency_map
+ A real-valued saliency-map that conveys regions used for
+ classification in the original sample.
+
+
+ Returns
+ -------
+ A numpy array of the same size as saliency_map with
+ the value False everywhere except at the positions inside
+ the bounding boxes, which will be True.
+ """
+ binary_mask = numpy.zeros_like(saliency_map, dtype=numpy.bool_)
+ for bbox in gt_bboxes:
+ binary_mask[
+ bbox.ymin : bbox.ymin + bbox.height,
+ bbox.xmin : bbox.xmin + bbox.width,
+ ] = True
+ return binary_mask
+
+
def _process_sample(
gt_bboxes: BoundingBoxes,
saliency_map: numpy.typing.NDArray[numpy.double],
@@ -291,13 +325,7 @@ def _process_sample(
# # Calculate localization metrics
# iou, ioda = _compute_max_iou_and_ioda(detected_box, gt_bboxes)
- # The binary_mask will be ON/True where the gt boxes are located
- binary_mask = numpy.zeros_like(saliency_map, dtype=numpy.bool_)
- for bbox in gt_bboxes:
- binary_mask[
- bbox.ymin : bbox.ymin + bbox.height,
- bbox.xmin : bbox.xmin + bbox.width,
- ] = True
+ binary_mask = _compute_binary_mask(gt_bboxes, saliency_map)
return (
# iou,
--
GitLab