MATLAB多边形容器（边数>>4）目标检测问题：大象检测失败排查

我尝试修改MATLAB代码以检测桌面图像中的大象，预期使用边数远多于4的多边形大致贴合大象轮廓标记目标区域，但多次尝试后，即使调用ginput(12)仍仅得到4边形，怀疑是参数设置存在问题。以下为初始代码及我修改后的代码：

初始代码 #

clear all
close all
clc
%% read images
% template
boxImage = imread('./immaginiObjectDetection/stapleRemover.jpg');
% desk
sceneImage = imread('./immaginiObjectDetection/clutteredDesk.jpg');
figure(1), clf, imshow(boxImage)
figure(2), clf, imshow(sceneImage)
% figure(1), clf, imagesc(boxImage)
% figure(2), clf, imagesc(sceneImage)
tic
%% keypoint detection
boxPoints = detectSURFFeatures(boxImage);
scenePoints = detectSURFFeatures(sceneImage);
figure(1), clf
imshow(boxImage), hold on
plot(selectStrongest(boxPoints,100)), hold off
figure(2), clf
imshow(sceneImage), hold on
plot(selectStrongest(scenePoints,100)), hold off
%% keypoint description
[boxFeatures, boxPoints]=extractFeatures(boxImage, boxPoints);
[sceneFeatures, scenePoints]=extractFeatures(sceneImage, scenePoints);
%% feature matching
boxPairs = matchFeatures(boxFeatures, sceneFeatures);
matchedBoxPoints = boxPoints(boxPairs(:,1),:);
matchedScenePoints = scenePoints(boxPairs(:,2),:);
showMatchedFeatures(boxImage, sceneImage, matchedBoxPoints, ...
    matchedScenePoints, 'montage');
%% geometric consistency check
[tform, inlierBoxPoints, inlierScenePoints]=...
    estimateGeometricTransform(matchedBoxPoints,...
    matchedScenePoints,'affine');
showMatchedFeatures(boxImage, sceneImage, inlierBoxPoints, ...
    inlierScenePoints, 'montage');
%% bounding box drawing
boxPoly = [1 1;
    size(boxImage,2) 1;
    size(boxImage,2) size(boxImage,1);
    1 size(boxImage,1);
    1 1];
newBoxPoly=transformPointsForward(tform,boxPoly);
figure, clf
imshow(sceneImage), hold on
line(newBoxPoly(:,1),newBoxPoly(:,2),'Color','y')
hold off
toc
%% more precise bounding box
figure, clf
imshow(boxImage)
[x,y]=ginput(4);
%%
x=[x; x(1)];
y=[y; y(1)];
newBoxPoly=transformPointsForward(tform,[x y]);
figure, clf
imshow(sceneImage), hold on
line(newBoxPoly(:,1),newBoxPoly(:,2),'Color','y')
hold off
toc

修改后的代码 #

clear all
close all
clc
%% read images
% template
boxImage = imread('./immaginiObjectDetection/elephant.jpg');
% desk
sceneImage = imread('./immaginiObjectDetection/clutteredDesk.jpg');
figure(1), clf, imshow(boxImage)
figure(2), clf, imshow(sceneImage)
% figure(1), clf, imagesc(boxImage)
% figure(2), clf, imagesc(sceneImage)
tic
%% keypoint detection
boxPoints = detectSURFFeatures(boxImage, 'MetricThreshold', 1000);
scenePoints = detectSURFFeatures(sceneImage);
figure(1), clf
imshow(boxImage), hold on
plot(selectStrongest(boxPoints,100)), hold off
figure(2), clf
imshow(sceneImage), hold on
plot(selectStrongest(scenePoints,100)), hold off
%% keypoint description
[boxFeatures, boxPoints]=extractFeatures(boxImage, boxPoints);
[sceneFeatures, scenePoints]=extractFeatures(sceneImage, scenePoints);
%% feature matching
boxPairs = matchFeatures(boxFeatures, sceneFeatures, 'MatchThreshold', 30, ...
    'Method','Exhaustive');
matchedBoxPoints = boxPoints(boxPairs(:,1),:);
matchedScenePoints = scenePoints(boxPairs(:,2),:);
showMatchedFeatures(boxImage, sceneImage, matchedBoxPoints, ...
    matchedScenePoints, 'montage');
%% geometric consistency check
[tform, inlierBoxPoints, inlierScenePoints]=...
    estimateGeometricTransform(matchedBoxPoints,...
    matchedScenePoints,'projective', 'Confidence', 90, 'MaxDistance', 3);
showMatchedFeatures(boxImage, sceneImage, inlierBoxPoints, ...
    inlierScenePoints, 'montage');
%% bounding box drawing
boxPoly = [1 1;
    size(boxImage,2) 1;
    size(boxImage,2) size(boxImage,1);
    1 size(boxImage,1);
    1 1];
newBoxPoly=transformPointsForward(tform,boxPoly);
figure, clf
imshow(sceneImage), hold on
line(newBoxPoly(:,1),newBoxPoly(:,2),'Color','y')
hold off
toc
%% more precise bounding box
figure, clf
imshow(boxImage)
[x,y]=ginput(12);
%%
x=[x; x(1)];
y=[y; y(1)];
newBoxPoly=transformPointsForward(tform,[x y]);
figure, clf
imshow(sceneImage), hold on
line(newBoxPoly(:,1),newBoxPoly(:,2),'Color','y')
hold off
toc

内容的提问来源于stack exchange，提问作者Francesco Scipione