It’s been a long time since I wrote a blog. In fact, the main thing is that I have nothing to write about. . . The level is limited, hey. . . . I'm currently working on robotics kinematics, so deep vision is negligent here.
However, I recently saw a previous article and found it very interesting, so I want to share it with you.
The name of the article is the title "Sliding Shapes for 3D Object Detection in Depth Images", which means to use sliding windows in depth images for object recognition.
The effect is as follows:
Next is the process:
I don't know why the words are so small, but I should be able to see it clearly. The process is to first train the CAD model, then render the depth map, extract the depth map features, and use it for SVM training, and finally use the sliding window for object recognition. The overall process is actually the same as HOG for pedestrian recognition. It is nothing more than a three-dimensional space. So the biggest difference here is what the extracted features are. The features are as follows:
His descriptor combines the point cloud, normal direction, shape and TSDF information, total synthesis.
The above is a schematic diagram of the sliding window and the training model, which still looks very simple.
This is the result of the example after the entire project is run, and the effect can be seen. I uploaded the example, you can download it here .