Object detection using ORB [closed]

asked 2020-05-26 10:32:39 -0500

albeh95 gravatar image

Hello everyone, I have a video given (video.avi) and a set of images of books to be detected inside the video. My code actually doesn't detect the right book, but it always seems to detect the same book without noticing differences between all of them, and also doesn't draw the rectangle around the book (not even around the only one that always detect). I'm posting my code and a screenshot of the result I get:

#include <opencv2/core.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/opencv.hpp>
#include <iostream>
#include <opencv2/core/utils/filesystem.hpp>
#include <opencv2/core/types_c.h>

using namespace std;
using namespace cv;

int main() {

vector<Mat> data;
vector<string> fn;
glob("C:/Users/albma/Desktop/Università/Computer Vision/labs/lab6/Lab 6 data/objects/*.png", fn, true);
for (size_t k = 0; k < fn.size(); ++k)
    Mat im = imread(fn[k]);
    if (im.empty()) continue; //only proceed if sucsessful
cvtColor(data[0], data[0], COLOR_BGR2GRAY);
Mat img_object = data[0].clone();

Ptr<FeatureDetector> detector;
vector<KeyPoint> keypoints_object, keypoints_scene;
Mat descriptors_object;
Mat descriptors_scene;

VideoCapture cap = VideoCapture("video.mov");
if (cap.isOpened()) {
    Mat img_scene;
    cap >> img_scene;
    cvtColor(img_scene, img_scene, COLOR_BGR2GRAY);
    Ptr<FeatureDetector> detector = ORB::create();
    Ptr<DescriptorExtractor> descriptor = ORB::create();

    detector->detect(img_object, keypoints_object);
    detector->detect(img_scene, keypoints_scene);

    descriptor->compute(img_object, keypoints_object, descriptors_object);
    descriptor->compute(img_object, keypoints_scene, descriptors_scene);

    cv::Ptr<cv::BFMatcher> matcher = cv::BFMatcher::create(cv::NORM_HAMMING);
    vector< std::vector<DMatch> > knn_matches;
    matcher->knnMatch(descriptors_object, descriptors_scene, knn_matches, 2);

    //-- Filter matches using the Lowe's ratio test
    const float ratio_thresh = 1.75f;
    std::vector<DMatch> good_matches;
    for (size_t i = 0; i < knn_matches.size(); i++)
        if (knn_matches[i][0].distance < ratio_thresh * knn_matches[i][1].distance)
    //-- Draw matches
    Mat img_matches;
    drawMatches(img_object, keypoints_object, img_scene, keypoints_scene, good_matches, img_matches, Scalar::all(-1),
        Scalar::all(-1), std::vector<char>(), DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);
    //-- Localize the object
    std::vector<Point2f> obj;
    std::vector<Point2f> scene;
    for (size_t i = 0; i < good_matches.size(); i++)
        //-- Get the keypoints from the good matches
    Mat H = findHomography(obj, scene, RANSAC);
    //-- Get the corners from the image_1 ( the object to be "detected" )
    std::vector<Point2f> obj_corners(4);
    obj_corners[0] = Point2f(0, 0);
    obj_corners[1] = Point2f((float)img_object.cols, 0);
    obj_corners[2] = Point2f((float)img_object.cols, (float)img_object.rows);
    obj_corners[3] = Point2f(0, (float)img_object.rows);
    std::vector<Point2f> scene_corners(4);

    perspectiveTransform(obj_corners, scene_corners, H);
    //-- Draw lines between the corners (the mapped object in the scene - image_2 )
    line(img_matches, scene_corners[0] + Point2f((float)img_object.cols, 0),
        scene_corners[1] + Point2f((float)img_object.cols, 0), Scalar(0, 255, 0), 4);
    line(img_matches, scene_corners[1] + Point2f((float)img_object.cols, 0),
        scene_corners[2] + Point2f((float)img_object.cols, 0), Scalar(0, 255, 0), 4);
    line(img_matches, scene_corners[2] + Point2f((float)img_object.cols, 0),
        scene_corners[3] + Point2f((float)img_object.cols, 0), Scalar(0, 255, 0), 4);
    line(img_matches, scene_corners[3] + Point2f((float)img_object.cols, 0),
        scene_corners[0] + Point2f((float)img_object.cols, 0), Scalar(0, 255, 0), 4);
    //-- Show detected matches
    resize(img_matches, img_matches ...
edit retag flag offensive reopen merge delete

Closed for the following reason the question is answered, right answer was accepted by albeh95
close date 2020-05-29 12:08:41.078800


You should have one template pic not 4.

supra56 gravatar imagesupra56 ( 2020-05-26 21:02:53 -0500 )edit

My code actually doesn't detect the right book,

that is expected. you cannot do real "object detection" like this, waste of time.

opencv's feature matching algorithms are meant to find a homography for parts of the same, known scene. they do not know anything about "objects", and you cannot use arbitrary images.

berak gravatar imageberak ( 2020-05-27 01:34:20 -0500 )edit

if you think of it, what you're trying is even to recognize certain books at the same time.

and detecting books is also terribly difficult, since the "inner class variance" is so high

berak gravatar imageberak ( 2020-05-27 01:49:23 -0500 )edit

So is there any way to solve the problem? Following some tutorials I've seen they use SIFT features, but I'm not allowed to do it. Also the whole algorithm looks the same of mine, so I really don't get it

albeh95 gravatar imagealbeh95 ( 2020-05-27 08:35:38 -0500 )edit

idk, what tutorials you've seen there, but you (or someone there) must have misunderstood the idea.

berak gravatar imageberak ( 2020-05-27 08:43:40 -0500 )edit

Since other people had the same problems (everyone ahahah) we asked to use SIFT features and now it works. Thanks y'all!

albeh95 gravatar imagealbeh95 ( 2020-05-27 16:16:44 -0500 )edit