Millions of people watched from the coffee shop monitoring, and Musk exclaimed that it was terrible! You drink coffee for a few minutes, AI knows everything

Article source: Xinzhiyuan

EDIT: Aeneas is so sleepy

In this coffee shop, how long the customer stayed and how many cups of coffee the clerk made is clearly visible under the AI camera. Millions of netizens watched and said it was terrible, and Musk was shocked.

We live in a world where there is less and less privacy.

Today, this video that went out on the Internet scared many people.

In a coffee shop, each customer stayed in the store for a few minutes, and how many cups of coffee each waiter brought to the customer, all of which are clearly shown in the video!

The video has only been released for more than ten hours, and more than 1 million netizens have watched it.

The netizen who posted the video said: This concept shows how coffee shops use AI to analyze baristas and customers. Please fully “enjoy” your privacy at the coffee shop. 😂

Another netizen said it was no surprise. As a consumer, you should know that many stores know everything about you the moment you enter.

The “Cambridge Analytica incident” pales in comparison.

(In 2018, Facebook admitted that the British data analysis company illegally obtained the information of 50 million Facebook users in 2016, and used this information to build a software program that predicted and influenced the results of the ballot box and successfully helped Trump won the presidential election.)

Even Musk himself appeared in the comment area, leaving two exclamation marks in a row.

If you think having an AI spying on staff and customers in a coffee shop is scary enough, the reality is that, if cost is not an issue, there can be thousands of drones in the sky sending real-time tracking data to regulators, and everything will be tracked and recorded.

You don’t even need a professional department, anyone can do tracking on an amateur drone, because current object detection and image recognition technology is just too powerful.

You know, when running 1080p streaming on a discrete graphics card a few years ago, the maximum capacity was only 6 objects.

The ubiquitous “eyes”

The reality is that our current world is full of cameras.

Among them, many companies have deployed very covert strategies to track consumers, all done through AI and visual recognition on video feeds.

For example, in Wal-Mart’s smart retail laboratory, IRL sensors and cameras allow staff to know everything in the store.

Fast food restaurants have also adopted AI technology for employee supervision. It is stipulated that employees must wear masks. If anyone takes off the mask, the manager will know immediately.

Also, our mobile location data is for sale.

Almost all mobile phone operators are selling data anonymously to retail stores, which can be said to be part of their core business.

Just Google “operator name + crowd insights” and the results will surprise you.

“Do you want to know how many people pass by a certain place in a certain period of time? What is their age and income status, and how many of them can become potential customers?”

Of course, the “crowd insight” service will emphasize that the data is anonymous, and the way the receipt is collected does not reveal personal privacy.

Someone said: Now that my data is collected, can I ask the company to pay me?

Regarding the cameras used in enterprises, some people in the comment area showed their own experiences——

“I work on the backend security camera system at the stadium, and what we release to the public is only 1/3 of the actual data.”

“It’s almost like in a movie, input your own face, and the system will recognize where you are.”

And to do all that, you just take any camera, install a $300 piece of software, and run until you run out of disk space.

** Pros and cons? **

In this regard, AI consulting expert Diego San Esteban shared his views:

He believes that AI monitoring certainly has many advantages, such as continuous monitoring of employee performance and productivity, allowing managers to better formulate strategies.

In addition, AI can also provide objective performance data to avoid human bias in evaluation.

There are also many shortcomings. The most criticized one is the violation of employees’ privacy rights, and it will also create an atmosphere of distrust in the enterprise, which will affect morale and job satisfaction.

AI also cannot adequately understand the context in which work is being done, and lacks human empathy.

And, it is likely to make mistakes, subject to the inherent bias of the training data, which is extremely unfair to employees.

Target detection algorithm

In fact, behind this controversial incident is a very common AI technology-target detection.

For example, given a photo of a city street, an object detection model will return a list of annotations or labels for all the different objects in the image: traffic lights, vehicles, road signs, buildings, etc.

These labels will contain the appropriate class for each object, such as “person”, and a “bounding box”, the rectangular area that completely encloses the object.

### Industry Application

Object detection is a critical task for humans: when entering a new room or scene, our first instinct is to visually assess the objects and people in it and then understand them.

Similar to humans, object detection plays a vital role in enabling computers to understand and interact with the visual world, and has been widely used in many industries:

- Site Security:

Object detection models can help improve workplace safety and security. For example, they can detect the presence of suspicious individuals or vehicles in sensitive areas. Even more creatively, it can ensure workers use personal protective equipment (PPE) such as gloves, helmets or masks.

- social media:

Object detection models can help identify the presence of a particular brand, product, logo, or person in digital media. Advertisers can use this information to gather data and show users more relevant ads. It also helps automate the process of detecting and flagging inappropriate or banned content.

- QC:

Object detection models enable automated review of visual data. Computers and cameras can analyze data in real time, automatically detecting and processing visual information and understanding its significance, reducing human intervention in tasks that require constant visual review. This is especially useful in manufacturing production quality control. Not only does it improve efficiency, it can also detect production anomalies that the human eye might miss, preventing potential production disruptions or product recalls.

Achieved 66 AP for the first time, the strongest SOTA algorithm dominates the list

Currently, in terms of the performance of the target detection algorithm, “DETRs with Collaborative Hybrid Assignments Training” from the domestic team dominates COCO with a score of 66 AP. This work has been accepted by ICCV 2023.

In the paper, the authors propose a novel collaborative hybrid assignment training scheme - Co-DETR, which can learn more efficient and effective DETR-based detectors from diverse label assignments.

By training multiple parallel auxiliary heads (supervised by one-to-many label assignment, such as ATSS and Faster RCNN), the new Co-DETR can easily boost the learning ability of encoder in end-to-end detector.

By extracting positive coordinates from these auxiliary heads for additional customized positive queries, Co-DETR can also improve the training efficiency of positive samples in the decoder.

Moreover, these auxiliary heads are discarded during inference, so the method does not introduce additional parameters and computational cost to the original detector, and also does not require manual non-maximum suppression (NMS).

Paper address:

project address:

- Encoder optimization:

The training scheme can easily boost the learning ability of the encoder in the end-to-end detector by training multiple parallel auxiliary heads supervised by one-to-many label assignment.

- Codec optimization:

Attention learning in the decoder is improved by extracting positive coordinates from these auxiliary heads for additional custom positive queries.

- SOTA performance:

Co-DETR equipped with ViT-L (304M parameters) is the first model to achieve 66.0% AP on COCO test-dev.

The experimental results show that based on the Swin-L backbone network, the Co-DETR method can improve the performance of the existing SOTA model DINO-Deformable-DETR from 58.5% to 59.5% (on the COCO validation set).

With the support of ViT-L backbone network, Co-DETR achieves 66.0% AP on COCO test-dev and 67.9% AP on LVIS validation set.

In addition, Co-DETR also achieves better performance with a smaller model size than previous methods.

References: