Deep Learning Based Pose Landmark

This guide provides instructions on how to utilize the provided Pose Landmark API within your application to identify pose landmarks in a given image.

Prerequisites

Ensure the following prerequisites are satisfied: Include the necessary headers mv_pose_landmark.h in your project #include <mv_pose_landmark.h>

Optionally, include additional headers for handling image decoding image_util.h or acquiring preview images from cameras #include <image_util.h> // Optional: Image decoding support #include <camera.h> // Optional: Acquiring preview images from cameras

Detect poses in an image

Follow these steps to implement pose landmark in your application:

Step 1: Initialize and Prepare First, create a pose landmark handle and prepare the environment for pose landmark:

mv_pose_landmark_h handle;
int ret = 0;

ret = mv_pose_landmark_create(&handle);
if (ret != MEDIA_VISION_ERROR_NONE) {
    // handle an error.
}

 ret = mv_pose_landmark_configure(handle);
 if (ret != MEDIA_VISION_ERROR_NONE) {
     // handle an error.
 }

ret = mv_pose_landmark_prepare(handle);
if (ret != MEDIA_VISION_ERROR_NONE) {
    // handle an error.
}

Step 2: Input Source Setup Next, set up the input source containing the image data. Here, we’ll demonstrate decoding an image file and filling the resulting data into a mv_source:

char filePath[1024];
unsigned char *dataBuffer = NULL;
size_t bufferSize = 0;
unsigned int width = 0;
unsigned int height = 0;
image_util_decode_h imageDecoder = NULL;
mv_source_h mv_source = NULL;
image_util_image_h decodedImage = NULL;

ret = mv_create_source(&mv_source);
if (ret != MEDIA_VISION_ERROR_NONE) {
    // handle an error.
}

ret = image_util_decode_create(&imageDecoder);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

/* Decode image and fill the image data to mv_source handle */
snprintf(filePath, 1024, "/path/to/pose_image.jpg");
ret = image_util_decode_set_input_path(imageDecoder, filePath);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

ret = image_util_decode_set_colorspace(imageDecoder, IMAGE_UTIL_COLORSPACE_RGB888);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

ret = image_util_decode_run2(imageDecoder, &decodedImage);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

ret = image_util_get_image(decodedImage, &width, &height, NULL, &dataBuffer, &bufferSize);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

ret = mv_source_fill_by_buffer(mv_source, dataBuffer, (unsigned int)bufferSize,
                              width, height, MEDIA_VISION_COLORSPACE_RGB888);
if (ret != MEDIA_VISION_ERROR_NONE) {
   free(dataBuffer);
   // handle an error.
}

ret = image_util_decode_destroy(imageDecoder);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

ret = image_util_destroy_image(decodedImage);
if (ret != IMAGE_UTIL_ERROR_NONE) {
    // handle an error.
}

Step 3: Pose Landmark Execution There are two modes available for executing pose landmark: synchronous and asynchronous. Choose the appropriate mode based on your application requirements:

[Synchronous Mode] For synchronous processing, call mv_pose_landmark_inference() with the prepared mv_source. This API will be returned after the completion of the inference request: // Detect poses in a given image. ret = mv_pose_landmark_inference(handle, mv_source); if (ret != MEDIA_VISION_ERROR_NONE) { // handle an error. }

Afterwards, retrieve the number of detected poses and obtain their respective bounding boxes: ``` unsigned long frame_number; unsigned int number_of_landmarks;

	ret = mv_pose_landmark_get_result_count(handle, &frame_number, &number_of_landmarks);
	if (ret!= MEDIA_VISION_ERROR_NONE) {
		// handle an error.
	}

	for (unsigned int idx = 0; idx < number_of_landmarks; ++idx) {
		unsigned int pos_x, pos_y;

		int ret = mv_pose_landmark_get_position(handle, idx, &pos_x, &pos_y);
		if (ret!= MEDIA_VISION_ERROR_NONE) {
			// handle an error.
		}
	}

	// Process position information...
    ```

[Asynchronous Mode] The asynchronous API, mv_pose_landmark_inference_async(), returns immediately after being called, and inference results can be obtained by creating a thread and calling the mv_pose_landmark_get_result_count() API within its callback function. If asynchronous processing is preferred, invoke mv_pose_landmark_inference_async() with the prepared mv_source, and handle the results via a callback function: ``` void pose_landmark_callback(void *user_data) { mv_pose_landmark_h handle = (mv_pose_landmark_h)user_data; bool is_loop_exit = false;

	while (!is_loop_exit) {
		unsigned long frame_number;
		unsigned int number_of_landmarks;

		int ret = mv_pose_landmark_get_result_count(handle, &frame_number, &number_of_landmarks);
		if (ret!= MEDIA_VISION_ERROR_NONE) {
			// handle an error.
		}

		for (unsigned int idx = 0; idx < number_of_landmarks; ++idx) {
			unsigned int pos_x, pos_y;

			int ret = mv_pose_landmark_get_position(handle, idx, &pos_x, &pos_y);
			if (ret!= MEDIA_VISION_ERROR_NONE) {
				// handle an error.
			}
		}
	}
}

void some_function()
{
	...

	// Detect poses in a given image.
	ret = mv_pose_landmark_inference_async(handle, mv_source);
	if (ret!= MEDIA_VISION_ERROR_NONE) {
		// handle an error.
	}

	// Create a new thread to wait for the inference result.
	thread *thread_handle = new thread(&pose_landmark_callback, (void *)handle);
	if (thread_handle == NULL) {
		// handle an error.
	}

	thread_handle->join();
}
```

Step 4: Cleanup Finally, clean up by releasing the allocated resources and destroying the handle: ret = mv_pose_landmark_destroy(handle); if (ret != MEDIA_VISION_ERROR_NONE) { // handle an error. }

  • Dependencies
    • Tizen 9.0 and Higher for TV
Edit this page