Update pspnet config (#3144)

* update pspnet cofig

* update readme

* update crop_padding

* rename postprocessing

Author: Julia Kamelina

models/public/pspnet-pytorch/README.md

@@ -17,7 +17,7 @@
 
 | Metric    | Value |
 | --------- | ----- |
-| mean_iou  | 70.6% |
+| mean_iou  | 70.1% |
 
 Accuracy metrics were obtained with fixed input resolution 512x512.
 

models/public/pspnet-pytorch/accuracy-check.yml

@@ -7,22 +7,20 @@ models:
       - name: VOC2012_Segmentation
         preprocessing:
           - type: resize
-            aspect_ratio_scale: frcnn_keep_aspect_ratio
-            dst_height: 512
-            dst_width: 2048
-          - type: crop
+            aspect_ratio_scale: fit_to_window
             size: 512
+          - type: padding
+            size: 512
+            pad_type: right_bottom
+            pad_value: '123.675, 116.28, 103.53'
         postprocessing:
+          - type: crop_padded_prediction
           - type: resize_segmentation_mask
-            apply_to: annotation
-            to_dst_image_size: True
+            apply_to: prediction
           - type: encode_segmentation_mask
             apply_to: annotation
-          - type: crop_segmentation_mask
-            size: 512
-            apply_to: annotation
         metrics:
           - type: mean_iou
             use_argmax: false
             presenter: print_scalar
-            reference: 0.706
+            reference: 0.701

tools/accuracy_checker/openvino/tools/accuracy_checker/postprocessor/README.md

@@ -64,6 +64,7 @@ Accuracy Checker supports following set of postprocessors:
     You can also use `size` instead in case when destination sizes are equal for all three dimensions.
 * `crop_or_pad-segmentation_mask` - performs central cropping if original mask size greater then destination size and padding in case, when source size lower than destination. Padding filling value is 0, realization - right-bottom.
   * `dst_width` and `dst_height` are destination width and height for keypoints resizing respectively. You can also use `size` instead in case when destination sizes are equal.  Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
+* `crop_padded_prediction` - performs cropping if original image was padded, to the padding size. Supported representations: `SegmentationAnotation`, `SegmentationPrediction`.
 * `heatmap2keypoints` - extract landmark keypoints from the heatmap. Supported representations: `FacialLandmarksHeatMapAnnotation`, `FacialLandmarksHeatMapPrediction`.
 * `clip_segmentation_mask` - clipping segmentation mask values. Supported representations: `BrainTumorSegmentationAnnotation`, `BrainTumorSegmentationPrediction`.
   * `min_value` - lower bound of range.

tools/accuracy_checker/openvino/tools/accuracy_checker/postprocessor/__init__.py

@@ -40,7 +40,7 @@
 from .clip_points import ClipPoints
 from .extend_segmentation_mask import ExtendSegmentationMask
 from .zoom_segmentation_mask import ZoomSegMask
-from .crop_segmentation_mask import CropSegmentationMask, CropOrPadSegmentationMask
+from .crop_segmentation_mask import CropSegmentationMask, CropOrPadSegmentationMask, CropPaddingSegmentationMask
 from .clip_segmentation_mask import ClipSegmentationMask
 from .normalize_boxes import NormalizeBoxes
 from .brats_postprocessing import SegmentationPredictionResample, TransformBratsPrediction
@@ -99,6 +99,7 @@
     'ZoomSegMask',
     'CropSegmentationMask',
     'CropOrPadSegmentationMask',
+    'CropPaddingSegmentationMask',
     'ClipSegmentationMask',
     'ArgMaxSegmentationMask',
 

tools/accuracy_checker/openvino/tools/accuracy_checker/postprocessor/crop_segmentation_mask.py

@@ -145,3 +145,31 @@ def process_mask(self, mask):
             cropped, offset_pad_height, offset_pad_width, self.dst_height, self.dst_width
         )
         return resized
+
+
+class CropPaddingSegmentationMask(Postprocessor):
+    __provider__ = 'crop_padded_prediction'
+
+    annotation_types = (SegmentationAnnotation, )
+    prediction_types = (SegmentationPrediction, )
+
+    def process_image(self, annotation, prediction):
+        raise RuntimeError("Since `process_image_with_metadata` is overridden, this method MUST NOT be called")
+
+    def process_image_with_metadata(self, annotation, prediction, image_metadata=None):
+        assert image_metadata and 'padding' in image_metadata, (
+            "Postprocessing step `crop_padded_prediction` cannot work without metadata with `padding` field")
+
+        top, left, bottom, right = image_metadata.get('padding', (0, 0, 0, 0))
+        for pred in prediction:
+            mask = pred.mask
+            if mask.ndim == 2:
+                h, w = mask.shape
+                mask = mask[top:(h - bottom), left:(w - right)]
+                pred.mask = mask
+                continue
+            _, h, w = mask.shape
+            mask = mask[:, top:(h - bottom), left:(w - right)]
+            pred.mask = mask
+
+        return annotation, prediction