OpenCV SURF目标检测技术咨询：角点变换与识别性能边界问题

Hey there, let's tackle your two questions about the unexpected bounding polygon issue and SURF detection limits:

1. Why add Point2f(img_object.cols, 0) to the corner points? #

First off, I think you're referring to defining the object's reference corner points (not modifying scene_corners directly). Let me break it down clearly:

When using Homography to map your target object to the scene, we need a full set of 4 reference points that define the complete bounding box of your original object image. These points are:

• (0, 0): Top-left corner of the object
• Point2f(img_object.cols, 0): Top-right corner of the object
• Point2f(img_object.cols, img_object.rows): Bottom-right corner
• Point2f(0, img_object.rows): Bottom-left corner

This full set tells the Homography matrix exactly what region of the object to warp to its position in the scene. If you skip or incorrectly modify any of these points (like omitting the top-right one), the transformed scene_corners will not form the correct polygon around your target—this is almost certainly why your bounding shape is behaving unexpectedly.

The Homography calculation relies on these corresponding points to warp the object's boundary accurately. Without the complete 4 corners, the warp will be skewed or incomplete.

2. What range of target orientation and distance works for effective SURF detection? #

SURF's effectiveness depends on how well it can match local features between your object and the scene. Here's a practical breakdown of reliable limits:

Orientation Range #

• SURF has built-in rotation invariance, but it's not unlimited. For stable detection, keep target rotation within ±60° relative to the original object image.
• Beyond ±90°, feature match counts drop sharply, and the Homography calculation may fail (or produce a wrong bounding box). Symmetric targets might work at 180°, but this is inconsistent and not recommended for reliable results.

Distance Range #

This ties directly to how large your target appears in the scene image:

• Minimum size: Your target should be at least ~30x30 pixels in the scene. Any smaller, and SURF can't extract enough distinct local features to match reliably.
• Maximum size: Avoid having the target fill the entire frame or be so large that its edges are cut off. Extreme over-sizing can lead to feature saturation or loss of edge details critical for matching.
• Scale variation: SURF handles moderate scale changes (up to ~3-5x difference in object size between the reference and scene) well. Beyond that, feature descriptors become too dissimilar to match accurately.

Also, quick side notes:

• Partial occlusion (more than 30% of the target covered) will drastically reduce detection accuracy.
• Extreme lighting changes can also throw off feature matching, even if orientation/distance is within range.

内容的提问来源于stack exchange，提问作者Komms