CNN multi-class object detection (Pascal VOC)

samples/cnn_voc.c

@@ -0,0 +1,109 @@
+/*
+ * Programming introduction with the SOD Embedded Convolutional/Recurrent Neural Networks (CNN/RNN) API.
+ * Copyright (C) PixLab | Symisc Systems, https://sod.pixlab.io
+ */
+/*
+* Compile this file together with the SOD embedded source code to generate
+* the executable. For example:
+*
+*  gcc sod.c cnn_voc.c -lm -Ofast -march=native -Wall -std=c99 -o sod_cnn_intro
+*  
+* Under Microsoft Visual Studio (>= 2015), just drop `sod.c` and its accompanying
+* header files on your source tree and you're done. If you have any trouble
+* integrating SOD in your project, please submit a support request at:
+* https://sod.pixlab.io/support.html
+*/
+/*
+* This simple program is a quick introduction on how to embed and start
+* experimenting with SOD without having to do a lot of tedious
+* reading and configuration.
+*
+* Make sure you have the latest release of SOD from:
+*  https://pixlab.io/downloads
+* The SOD Embedded C/C++ documentation is available at:
+*  https://sod.pixlab.io/api.html
+*/
+#include <stdio.h>
+#include "sod.h"
+int main(int argc, char *argv[])
+{
+	/* Input image (pass a path or use the test image shipped with the samples ZIP archive) */
+	const char *zInput = argc > 1 ? argv[1] : "./test.png";
+	/* Draw detection boxes (i.e. rectangles) on this output image which
+	 * is a copy of the input plus the boxes.
+	 */
+	const char *zOut = argc > 2 ? argv[2] : "./out.png";
+	/*
+	 * The CNN handle that should perform the detection process */
+	sod_cnn *pNet;
+	/* Load the input image */
+	sod_img imgIn = sod_img_load_from_file(zInput,SOD_IMG_COLOR/* Full colors*/);
+	if (imgIn.data == 0) {
+		/* Invalid path, unsupported format, memory failure, etc. */
+		puts("Cannot load input image..exiting");
+		return 0;
+	}
+	/* Make a copy so we can draw anything we want. */
+	sod_img imgOut = sod_copy_image(imgIn);
+	int rc;
+	const char *zErr; /* Error log if any */
+	/*
+	 * Create our CNN handle using the built-in fast
+	 * architecture trained on the Pascal VOC dataset
+	 * and is able to detect 20 classes of objects at
+	 * real-time on a modern CPU.
+	 */
+	rc = sod_cnn_create(&pNet, ":voc", "./tiny20.sod", &zErr);
+	/*
+	 * ":voc" is the magic word for the built-in Pascal VOC (20 classes) 
+	 * fast architecture. The list of built-in Magic words (pre-ready to use 
+	 * configurations and their associated models) are documented here:
+	 * https://sod.pixlab.io/c_api/sod_cnn_create.html.
+	 *
+	 * "tiny20.sod" is the pre-trained model associated with the ":fast" architecture 
+	 *  and is available to download from https://pixlab.io/downloads
+	 */
+	if (rc != SOD_OK) {
+		/* Display the error message and exit */
+		puts(zErr);
+		return 0;
+	}
+	/*
+	 * A sod_box instance always store the coordinates for each detected object
+	 * returned by the CNN via sod_cnn_predict() as we'll see later.
+	 */
+	sod_box *box;
+	int i, nbox;
+	/* Prepare our input image for the detection process which 
+	 * is resized to the network dimension (This op is always very fast)
+	 */
+	float * blob = sod_cnn_prepare_image(pNet, imgIn);
+	if (!blob) {
+		/* Very unlikely this happen: Invalid architecture, out-of-memory */
+		puts("Something went wrong while preparing image..");
+		return 0;
+	}
+	puts("Starting CNN object detection");
+	/* Detect.. */
+	sod_cnn_predict(pNet, blob, &box, &nbox);
+	/* Report the detection result. */
+	printf("%d object(s) were detected..\n",nbox);
+	for (i = 0; i < nbox; i++) {
+		/* Report the coordinates, name and score of the current detected object */
+		printf("(%s) X:%d Y:%d Width:%d Height:%d score:%f%%\n", box[i].zName, box[i].x, box[i].y, box[i].w, box[i].h, box[i].score * 100);
+		if( box[i].score < 0.3) continue;   /* Discard low score detection, remove if you want to report all objects */
+		/*
+		 * Draw a rose (RGB: 255,0,255) rectangle of width 3 on the object coordinates. */
+		sod_image_draw_bbox_width(imgOut, box[i], 3, 255., 0, 225.);
+		/* Of course, one could draw a circle via sod_image_draw_circle() or 
+		 * crop the entire region via sod_crop_image() instead of drawing a rectangle. */
+	}
+	/* Finally save our output image with the boxes drawn on it */
+	sod_img_save_as_png(imgOut, zOut);
+	/* Cleanup */
+	sod_free_image(imgIn);
+	sod_free_image(imgOut);
+	/* Release all resources allocated to the CNN handle */
+	sod_cnn_destroy(pNet);
+	return 0;
+}