18th sessions on FPGA machine learning

In the 18th lecture, we talked about machine vision. Processing video information, understanding the image, everyone has a mobile phone. Everyone has a camera. A two-dimensional graph restores a three-dimensional graph. Analyze the surface. Automatic inspection (aliens ). Understand human expressions. A story is a big problem. The illusion of human eyes. This is 3D perception, which can recognize near-distance and foreground background face recognition. You can find your face, but you don't know who it is. What's better at present is Tsinghua's Research Institute.

In the 18th lecture, we talked about machine vision.

Process video information and understand images

Everyone has a cell phone. Everyone has a camera.

A two-dimensional graph restores a three-dimensional graph. Analyze the surface. Automatic inspection (aliens ).

Understand human expressions.

A story is a big problem.

The illusion of human eyes.

This is 3D sensing, which can identify the distance and foreground

Face recognition. You can find your face, but you don't know who it is. What's best at present is the algorithm developed by Tsinghua Research Institute.

Judge the process.

Fingerprint recognition is now very good.

Currently, a company can perform 3D restoration.

3D reconstruction.

 

The same person, but too many changes. This is the difficulty of machine vision.

Here is an image rotation process.

The impact of light is also very significant for machine vision.

This is occlusion. Great impact. You must have a global understanding and a local understanding.

Emotion. Human expressions are acceptable. To identify animals, you also need to recognize the hold status and separate them.

The shape foreground background must be separated.

 

Machine vision is actually such a process, with many features in the middle, and different combinations are different things.

The key is how to extract features.

Features have the largest amount of information and need to be stable and not easily influenced by conditions such as light.

Can match the same image?

From this perspective, whether the extracted features are unaffected or not is well recognized.

This is the growth of people. The features that can be extracted by machines remain unchanged,

This is even harder.

A graph has a large scale, a small scale, and different angles. Then, find out which part of the graph has a large scale.

Recognition of the same book.

What is good information, or feature, is the last corner, because there is information change as long as there is slight change, which is easy to identify.

 

In fact, the method of 3D reconstruction of Avatar is to calibrate the expression changes from the green points on the face on the right. And then map back to the left.

Model: describes the reconstruction method.

What we hope is that different people go to the same location, but the angle is different and the light is different. It can be melted to construct a three-dimensional realistic graph.

Pedestrian detection.

 

I have limited capabilities, but I have studied hard. This is machine vision. The journey is far away.