Efficientdet

docs/structure.py

@@ -1,5 +1,6 @@
 from paz.backend import angles
 from paz.backend import boxes
+from paz.backend import anchors
 from paz.backend import camera
 from paz.backend import render
 from paz.backend import keypoints
@@ -58,11 +59,29 @@
             boxes.to_one_hot,
             boxes.to_normalized_coordinates,
             boxes.to_corner_form,
-            boxes.extract_bounding_box_corners
+            boxes.extract_bounding_box_corners,
+            boxes.scale_box
         ],
     },
 
 
+    {
+        'page': 'backend/anchors.md',
+        'functions': [
+            anchors.build_anchors,
+            anchors.build_octaves,
+            anchors.build_aspect,
+            anchors.build_scales,
+            anchors.build_strides,
+            anchors.make_branch_boxes,
+            anchors.compute_box_coordinates,
+            anchors.build_base_anchor,
+            anchors.compute_aspect_size,
+            anchors.compute_anchor_dims,
+            anchors.compute_anchor_centres
+        ],
+    },
+
     {
         'page': 'backend/keypoints.md',
         'functions': [
@@ -190,7 +209,8 @@
             image.get_rotation_matrix,
             image.calculate_image_center,
             image.get_affine_transform,
-            image.get_scaling_factor
+            image.get_scaling_factor,
+            image.scale_resize
         ],
     },
 
@@ -262,7 +282,15 @@
         'functions': [
             models.detection.SSD300,
             models.detection.SSD512,
-            models.detection.HaarCascadeDetector
+            models.detection.HaarCascadeDetector,
+            models.detection.EFFICIENTDETD0,
+            models.detection.EFFICIENTDETD1,
+            models.detection.EFFICIENTDETD2,
+            models.detection.EFFICIENTDETD3,
+            models.detection.EFFICIENTDETD4,
+            models.detection.EFFICIENTDETD5,
+            models.detection.EFFICIENTDETD6,
+            models.detection.EFFICIENTDETD7,
         ],
     },
 
@@ -395,7 +423,9 @@
             processors.ReplaceLowerThanThreshold,
             processors.GetNonZeroValues,
             processors.GetNonZeroArguments,
-            processors.FlipLeftRightImage
+            processors.FlipLeftRightImage,
+            processors.DivideStandardDeviationImage,
+            processors.ScaledResize
         ]
     },
 
@@ -454,7 +484,13 @@
             processors.NonMaximumSuppressionPerClass,
             processors.FilterBoxes,
             processors.OffsetBoxes2D,
-            processors.CropImage
+            processors.CropImage,
+            processors.BoxesToBoxes2D,
+            processors.BoxesWithOneHotVectorsToBoxes2D,
+            processors.BoxesWithClassArgToBoxes2D,
+            processors.RoundBoxes,
+            processors.RemoveClass,
+            processors.ScaleBox
         ]
     },
 
@@ -590,7 +626,9 @@
             pipelines.DetectSingleShot,
             pipelines.DetectHaarCascade,
             pipelines.SSD512HandDetection,
-            pipelines.SSD512MinimalHandPose
+            pipelines.SSD512MinimalHandPose,
+            pipelines.SSDPreprocess,
+            pipelines.SSDPostprocess
         ]
     },
 

examples/efficientdet/detection.py

@@ -1,17 +1,10 @@
 import numpy as np
-from paz import processors as pr
+import paz.processors as pr
 from paz.abstract import SequentialProcessor, Processor
-from paz.pipelines.detection import (DetectSingleShot, SSDPreprocess,
-                                     SSDPostprocess)
-from efficientdet import (EFFICIENTDETD0, EFFICIENTDETD1, EFFICIENTDETD2,
-                          EFFICIENTDETD3, EFFICIENTDETD4, EFFICIENTDETD5,
-                          EFFICIENTDETD6, EFFICIENTDETD7)
-from processors import (DivideStandardDeviationImage, ScaledResize, ScaleBox,
-                        NonMaximumSuppressionPerClass, FilterBoxes,
-                        ToBoxes2D, RemoveClass, RoundBoxes)
-
-B_IMAGENET_STDEV, G_IMAGENET_STDEV, R_IMAGENET_STDEV = 57.3, 57.1, 58.4
-RGB_IMAGENET_STDEV = (R_IMAGENET_STDEV, G_IMAGENET_STDEV, B_IMAGENET_STDEV)
+from paz.pipelines.detection import DetectSingleShot
+from paz.models import (EFFICIENTDETD0, EFFICIENTDETD1, EFFICIENTDETD2,
+                        EFFICIENTDETD3, EFFICIENTDETD4, EFFICIENTDETD5,
+                        EFFICIENTDETD6, EFFICIENTDETD7)
 
 
 class DetectSingleShotEfficientDet(Processor):
@@ -65,12 +58,12 @@ class EfficientDetPreprocess(SequentialProcessor):
             per channel on ImageNet.
     """
     def __init__(self, model, mean=pr.RGB_IMAGENET_MEAN,
-                 standard_deviation=RGB_IMAGENET_STDEV):
+                 standard_deviation=pr.RGB_IMAGENET_STDEV):
         super(EfficientDetPreprocess, self).__init__()
         self.add(pr.CastImage(float))
         self.add(pr.SubtractMeanImage(mean=mean))
-        self.add(DivideStandardDeviationImage(standard_deviation))
-        self.add(ScaledResize(image_size=model.input_shape[1]))
+        self.add(pr.DivideStandardDeviationImage(standard_deviation))
+        self.add(pr.ScaledResize(image_size=model.input_shape[1]))
 
 
 class EfficientDetPostprocess(Processor):
@@ -93,18 +86,17 @@ def __init__(self, model, class_names, score_thresh, nms_thresh,
         self.postprocess = pr.SequentialProcessor([
             pr.Squeeze(axis=None),
             pr.DecodeBoxes(model.prior_boxes, variances),
-            RemoveClass(class_names, class_arg)])
-        self.scale = ScaleBox()
-        self.nms_per_class = NonMaximumSuppressionPerClass(nms_thresh)
-        self.filter_boxes = FilterBoxes(score_thresh)
-        self.to_boxes2D = ToBoxes2D(class_names)
-        self.round_boxes = RoundBoxes()
-
-    def call(self, output, image_scales):
+            pr.RemoveClass(class_names, class_arg)])
+        self.scale = pr.ScaleBox()
+        self.nms_per_class = pr.NonMaximumSuppressionPerClass(
+            nms_thresh, conf_thresh=score_thresh)
+        self.to_boxes2D = pr.ToBoxes2D(class_names)
+        self.round_boxes = pr.RoundBoxes2D()
+
+    def call(self, output, image_scale):
         box_data = self.postprocess(output)
-        box_data = self.scale(box_data, image_scales)
-        box_data, class_data = self.nms_per_class(box_data)
-        box_data = self.filter_boxes(box_data, class_data)
+        box_data = self.scale(box_data, image_scale)
+        box_data = self.nms_per_class(box_data)
         boxes2D = self.to_boxes2D(box_data)
         boxes2D = self.round_boxes(boxes2D)
         return boxes2D
@@ -329,11 +321,8 @@ def __init__(self, score_thresh=0.60, nms_thresh=0.45, draw=True):
         names = get_class_names('VOC')
         model = EFFICIENTDETD0(num_classes=len(names),
                                base_weights='VOC', head_weights='VOC')
-        preprocess = SSDPreprocess(model)
-        postprocess = SSDPostprocess(
-            model, names, score_thresh, nms_thresh, class_arg=0)
         super(EFFICIENTDETD0VOC, self).__init__(
-             model, names, preprocess, postprocess, draw=draw)
+            model, names, score_thresh, nms_thresh, draw=draw)
 
 
 def process_outputs(outputs):