Image resize fix

[Beerwalker]

Fix for incorrect image resize. Image was not "letterboxed"(just resized with aspect ratio distortions) when using API function Detector::detect from yolo_v2_class.cpp.
Since resize is already performed inside Detector::detect, functions detect_resize and mat_to_image_resize were removed.

[AlexeyAB]

We shouldn't use letterbox_image() and resize_image() functions from the Darknet.

Fix for incorrect image resize. Image was not "letterboxed"(just resized with aspect ratio distortions) when using API function Detector::detect from yolo_v2_class.cpp.

1. There are pros and cons for each of 3 resize approaches: Resizing : keeping aspect ratio, or not #232 (comment)
2. For the most of my cases the resize() is better than letterbox()
3. Since we use resize() for Training in this repo, then we should use resize() for Detection too.

---

• functions resize_image() and letterbox_image() from Darknet are very slow, they are a bottleneck for performance, that is why I accelerated 4x times FPS for detection on video (FullHD/4K): What's the different between Darknet here and the official one? #529 (comment)
  by implementing resizing by using OpenCV cv::resize() function for the detector demo:

  darknet/src/image.c

  Line 1069 in a9fef1b

  cvResize(src, *in_img, CV_INTER_LINEAR);

Since resize is already performed inside Detector::detect, functions detect_resize and mat_to_image_resize were removed.

• this function cv::resize() is more than 4x times faster than resize_image() from Darknet:

  darknet/src/yolo_v2_class.hpp

  Lines 96 to 108 in a9fef1b

  	std::shared_ptr<image_t> mat_to_image_resize(cv::Mat mat) const
  	{
  	if (mat.data == NULL) return std::shared_ptr<image_t>(NULL);
  	cv::Size network_size = cv::Size(get_net_width(), get_net_height());
  	cv::Mat det_mat;
  	if (mat.size() != network_size)
  	cv::resize(mat, det_mat, network_size);
  	else
  	det_mat = mat; // only reference is copied
  	return mat_to_image(det_mat);
  	}

• after the image has been resized to the network size by using cv::resize(), there will not be called function resize_image(), just will be called memcpy() that has overhead less than 0.0...01%:

  darknet/src/yolo_v2_class.cpp

  Line 265 in a9fef1b

  memcpy(sized.data, im.data, im.w*im.h*im.c * sizeof(float));

[Beerwalker]

Thanks for a reply.

We shouldn't use letterbox_image() and resize_image() functions from the Darknet.

Ok, i accept that (however this approach isn't so great in my case). Have you considered using other interpolation methods for OpenCV cv::resize()? Since we usually downsample the common practice is to use CV_INTER_AREA for this task to obtain better image quality then CV_INTER_LINEAR. For upsampling CV_INTER_CUBIC is often used.
If you are interested in different interpolation methods for OpenCV you might want to look at this quick reference:
http://tanbakuchi.com/posts/comparison-of-openv-interpolation-algorithms/

[shubham-shahh]

We shouldn't use letterbox_image() and resize_image() functions from the Darknet.

Fix for incorrect image resize. Image was not "letterboxed"(just resized with aspect ratio distortions) when using API function Detector::detect from yolo_v2_class.cpp.

1. There are pros and cons for each of 3 resize approaches: #232 (comment)
2. For the most of my cases the resize() is better than letterbox()
3. Since we use resize() for Training in this repo, then we should use resize() for Detection too.

• functions resize_image() and letterbox_image() from Darknet are very slow, they are a bottleneck for performance, that is why I accelerated 4x times FPS for detection on video (FullHD/4K): #529 (comment)
  by implementing resizing by using OpenCV cv::resize() function for the detector demo:

  darknet/src/image.c

  Line 1069 in a9fef1b

  cvResize(src, *in_img, CV_INTER_LINEAR);

Since resize is already performed inside Detector::detect, functions detect_resize and mat_to_image_resize were removed.

• this function cv::resize() is more than 4x times faster than resize_image() from Darknet:

  darknet/src/yolo_v2_class.hpp

  Lines 96 to 108 in a9fef1b

  	std::shared_ptr<image_t> mat_to_image_resize(cv::Mat mat) const
  	{
  	if (mat.data == NULL) return std::shared_ptr<image_t>(NULL);
  	cv::Size network_size = cv::Size(get_net_width(), get_net_height());
  	cv::Mat det_mat;
  	if (mat.size() != network_size)
  	cv::resize(mat, det_mat, network_size);
  	else
  	det_mat = mat; // only reference is copied
  	return mat_to_image(det_mat);
  	}

• after the image has been resized to the network size by using cv::resize(), there will not be called function resize_image(), just will be called memcpy() that has overhead less than 0.0...01%:

  darknet/src/yolo_v2_class.cpp

  Line 265 in a9fef1b

  memcpy(sized.data, im.data, im.w*im.h*im.c * sizeof(float));

Does it resize the image with black paddings or just increase or decrease the height and width?