support flownet2 with arbitary input size

revise default proto to match the filename in documentations

fix a bug

beautify python codes

fix bug

beautify codes

add test samples with larger/smaller size

remove unless code

using bytearray without creating tmp file

remove useless codes

Author: zhaoyue-zephyrus

modules/dnn/src/layers/scale_layer.cpp

@@ -15,6 +15,7 @@ Implementation of Scale layer.
 #include "../op_inf_engine.hpp"
 #include "../ie_ngraph.hpp"
 
+#include <opencv2/imgproc.hpp>
 #include <opencv2/dnn/shape_utils.hpp>
 
 namespace cv
@@ -324,7 +325,7 @@ class DataAugmentationLayerImpl CV_FINAL : public DataAugmentationLayer
                          std::vector<MatShape> &internals) const CV_OVERRIDE
     {
         CV_Assert_N(inputs.size() == 1, blobs.size() == 3);
-        CV_Assert_N(blobs[0].total() == 1, blobs[1].total() == total(inputs[0], 1),
+        CV_Assert_N(blobs[0].total() == 1,
                     blobs[2].total() == inputs[0][1]);
 
         outputs.assign(1, inputs[0]);
@@ -347,15 +348,20 @@ class DataAugmentationLayerImpl CV_FINAL : public DataAugmentationLayer
         float* outData = outputs[0].ptr<float>();
 
         Mat data_mean_cpu = blobs[1].clone();
+        Mat mean_resize = Mat(inputs[0].size[3], inputs[0].size[2], CV_32FC3);
+        Mat mean_3d = Mat(data_mean_cpu.size[3], data_mean_cpu.size[2], CV_32FC3, data_mean_cpu.ptr<float>(0));
+        resize(mean_3d, mean_resize, Size(inputs[0].size[3], inputs[0].size[2]));
+        int new_size[] = {1, mean_resize.channels(), mean_resize.cols, mean_resize.rows};
+        Mat data_mean_cpu_resize = mean_resize.reshape(1, *new_size);
         Mat data_mean_per_channel_cpu = blobs[2].clone();
 
-        const int numWeights = data_mean_cpu.total();
+        const int numWeights = data_mean_cpu_resize.total();
         CV_Assert(numWeights != 0);
 
         ++num_iter;
         if (num_iter <= recompute_mean)
         {
-            data_mean_cpu *= (num_iter - 1);
+            data_mean_cpu_resize *= (num_iter - 1);
             const int batch = inputs[0].size[0];
             float alpha = 1.0 / batch;
 
@@ -364,15 +370,15 @@ class DataAugmentationLayerImpl CV_FINAL : public DataAugmentationLayer
                 Mat inpSlice(1, numWeights, CV_32F, inpData);
                 inpSlice = alpha * inpSlice;
 
-                add(data_mean_cpu.reshape(1, 1), inpSlice, data_mean_cpu.reshape(1, 1));
+                add(data_mean_cpu_resize.reshape(1, 1), inpSlice, data_mean_cpu_resize.reshape(1, 1));
                 inpData += numWeights;
             }
-            data_mean_cpu *= (1.0 / num_iter);
+            data_mean_cpu_resize *= (1.0 / num_iter);
 
-            int newsize[] = {blobs[1].size[1], (int)blobs[1].total(2)};
-            reduce(data_mean_cpu.reshape(1, 2, &newsize[0]), data_mean_per_channel_cpu, 1, REDUCE_SUM, CV_32F);
+            int newsize[] = {inputs[0].size[1], (int)inputs[0].total(2)};
+            reduce(data_mean_cpu_resize.reshape(1, 2, &newsize[0]), data_mean_per_channel_cpu, 1, REDUCE_SUM, CV_32F);
 
-            int area = blobs[1].total(2);
+            int area = inputs[0].total(2);
             data_mean_per_channel_cpu *= (1.0 / area);
         }
 
@@ -387,7 +393,7 @@ class DataAugmentationLayerImpl CV_FINAL : public DataAugmentationLayer
                 Mat inpSlice(1, numWeights, CV_32F, inpData);
                 Mat outSlice(1, numWeights, CV_32F, outData);
 
-                add(inpSlice, (-1) * data_mean_cpu, outSlice);
+                add(inpSlice, (-1) * data_mean_cpu_resize, outSlice);
                 inpData += numWeights;
                 outData += numWeights;
             }

modules/dnn/test/test_layers.cpp

@@ -646,6 +646,8 @@ TEST_P(Test_Caffe_layers, DataAugmentation)
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
     testLayerUsingCaffeModels("data_augmentation", true, false);
+    testLayerUsingCaffeModels("data_augmentation_2x1", true, false);
+    testLayerUsingCaffeModels("data_augmentation_8x6", true, false);
 }
 
 TEST_P(Test_Caffe_layers, Resample)

samples/dnn/optical_flow.py

@@ -5,7 +5,7 @@
 Original repo:  https://github.com/lmb-freiburg/flownet2.
 
 Download the converted .caffemodel model from https://drive.google.com/open?id=16qvE9VNmU39NttpZwZs81Ga8VYQJDaWZ
-and .prototxt from https://drive.google.com/open?id=19bo6SWU2p8ZKvjXqMKiCPdK8mghwDy9b.
+and .prototxt from https://drive.google.com/file/d/1RyNIUsan1ZOh2hpYIH36A-jofAvJlT6a/view?usp=sharing.
 Otherwise download original model from https://lmb.informatik.uni-freiburg.de/resources/binaries/flownet2/flownet2-models.tar.gz,
 convert .h5 model to .caffemodel and modify original .prototxt using .prototxt from link above.
 '''
@@ -18,7 +18,7 @@
 
 class OpticalFlow(object):
     def __init__(self, proto, model, height, width):
-        self.net = cv.dnn.readNet(proto, model)
+        self.net = cv.dnn.readNetFromCaffe(proto, model)
         self.net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
         self.height = height
         self.width = width
@@ -62,9 +62,9 @@ def motion_to_color(self, flow):
     parser = argparse.ArgumentParser(description='Use this script to calculate optical flow using FlowNetv2',
                                      formatter_class=argparse.ArgumentDefaultsHelpFormatter)
     parser.add_argument('-input', '-i', required=True, help='Path to input video file. Skip this argument to capture frames from a camera.')
-    parser.add_argument('--height', default=320, help='Input height')
-    parser.add_argument('--width',  default=448, help='Input width')
-    parser.add_argument('--proto', '-p', default='FlowNet2_deploy.prototxt', help='Path to prototxt.')
+    parser.add_argument('--height', default=320, type=int, help='Input height')
+    parser.add_argument('--width', default=448, type=int, help='Input width')
+    parser.add_argument('--proto', '-p', default='FlowNet2_deploy_anysize.prototxt', help='Path to prototxt.')
     parser.add_argument('--model', '-m', default='FlowNet2_weights.caffemodel', help='Path to caffemodel.')
     args, _ = parser.parse_known_args()
 
@@ -75,7 +75,25 @@ def motion_to_color(self, flow):
     cv.namedWindow(winName, cv.WINDOW_NORMAL)
     cap = cv.VideoCapture(args.input if args.input else 0)
     hasFrame, first_frame = cap.read()
-    opt_flow = OpticalFlow(args.proto, args.model, args.height, args.width)
+
+    divisor = 64.
+    var = {}
+    var['ADAPTED_WIDTH'] = int(np.ceil(args.width/divisor) * divisor)
+    var['ADAPTED_HEIGHT'] = int(np.ceil(args.height/divisor) * divisor)
+    var['SCALE_WIDTH'] = args.width / float(var['ADAPTED_WIDTH'])
+    var['SCALE_HEIGHT'] = args.height / float(var['ADAPTED_HEIGHT'])
+
+    config = ''
+    proto = open(args.proto).readlines()
+    for line in proto:
+        for key, value in var.items():
+            tag = "$%s$" % key
+            line = line.replace(tag, str(value))
+        config += line
+
+    caffemodel = open(args.model, 'rb').read()
+
+    opt_flow = OpticalFlow(bytearray(config.encode()), caffemodel, var['ADAPTED_HEIGHT'], var['ADAPTED_WIDTH'])
     while cv.waitKey(1) < 0:
         hasFrame, second_frame = cap.read()
         if not hasFrame: