How Data Annotation Services Fuel Self-Driving Industry 2021? — Part3
Unlike the clear answer to the battle between vision and laser, there is still no clear answer as to which is more important to automatic driving, algorithm, or computing power.
At present, it is difficult to say either the algorithmic group or the computing power group wins the game in a short time. Large automobile manufacturers prefer algorithmic groups because of lower costs and large vehicle output. More training data can better algorithm performance.
For example, Toyota just announced in May 2021 that the next generation of automatic driving models will adopt the technology jointly developed by Mobileye and ZF.
On the other hand, auto manufacturers who have developed an R&D team will undoubtedly choose the computing power, represented by NVIDIA, but it also brings another problem: computing power competition.
NIO has indeed prepared a NAD automatic driving system for ET7, which has a power of up to 1016TOPS. Compared with the 2.5TOPS chip of the previous model, it has increased by more than 400 times. But behind the soaring computing power, how can the four Orin chips work together and be used? NIO team did not announce the detailed plan.
What’s more, some auto companies which do not even have an automatic driving algorithm team at present, are willing to use high-power chips.
How much computing power is enough?
Such a top-level configuration of an automatic driving system requires 60+TOPS computing power to complete all data processing, and a 100TOPS computing chip is enough to deal with all data processing needs.
But this does not mean that high computing power has no value at all. For example, the shadow model adopted by Tesla and Huawei is a great computing power method.
Shadow mode is to let two sets of automatic driving software running at the same time, or when driving manually, there is always a set of automatic driving software running in the background. When the two decisions are different, or when the software decisions are different from the manual decisions, the system records all the data for algorithm upgrading and optimization.
At present, as there is a lack of automatic driving evaluation standards, car companies can only use the power figure, which is a simple and easy-to-understand indicator for consumers, to show the intelligent level of products.
There is also an important indicator — the number of manual takeovers per 100 km. However, this data also has a major defect, hard to be comparable. Similarly, in the 10000 km test, one vehicle is under complex road conditions in the urban area, and the other is on a freeway with few vehicles.
The current situation of automatic driving evaluation standard is that “there is no global authentic standard at present. A unified standard will be up in the end — the winner takes all. The player who will be in mass production and in really large-scale industrialization, its standard is the one that everyone will follow.
In 2004, DRAPA held the first million-dollar auto-driving challenge, which was the starting point of auto driving. At that time, the best team, the Carnegie Mellon college team, only making their self-driving vehicles go 10 kilometers in the desert.
Today, after 17 years, automatic driving technology has popularized.
Many people mentioned that automatic driving will enter the crucial stage of solving the long tail problem. In this critical stage, it is likely that some star companies will fall in the last mile.
For example, Waymo, Google’s star autopilot company, completed its first financing outside its parent company alpha in the first half of this year. In March and May, it raised more than $3 billion for two consecutive times. It seems to be good news, but considering that the valuation given by this financing is less than $40 billion, compared with the previous valuation peak of $175 billion, the capital market has returned from fanaticism to rationality.
The rapid development of automatic driving and the ultimate problem that has not yet found a solution. The road is blocked and long, but just out of the wild era.
Last year, at the CVPR 2019, Andrej Karpathy, the Senior Director of AI at Tesla responded to the question below:
how to estimate the volume of labeled data required to train and validate the self-driving cars for a particular scenario?”
378 hours of data.
The more accurate annotation is, the better algorithm performance will be.
Any tiny error during a driving experience may lead to dreadful results. Nowadays, people are more and more concerned about the driving safety issue as several self-driving automobile accidents happened.
For technology-type car companies, from 0 to 1 has been completed, but they still need to reduce costs and increase efficiency and move towards larger-scale production.
With the tremendous amount of training data and the high accuracy requirement, a high-quality data annotation service is crucial to guarantee autonomous vehicles are safe for the public.
Common Data Labeling Types Include:
It’s challenging for self-driving manufacturers to internally meet the burgeoning demand for high-quality data annotation.
The labeler used to annotate point by point, which cost lots of time.
3D annotation and video annotation are considered the toughest services in data labeling. At present, object tracking algorithms based on machine learning have already assisted video annotation. The annotator annotates the objects on the first frame, and then the algorithm tracks the ones in the subsequent frames. The annotator only needs to adjust the annotation when the algorithm doesn’t function well. It is 100 times faster than before.
Thanks to an AI-assisted system, the corresponding parts can be automatic can be transcribed and the human only needs to check and modify the wrong parts.
Nowadays, some AI-assisted tools come to practice, standing out in 2 factors.
- Cost reducing: With the help of AI-assisted capabilities, clients can save more money as the labor cost goes down.
- Time reducing: Make the large-scale requirement of training data done in a short time. Using AI-assisted tool can improve efficiency multiple times
Can we get rid of the human workforce?
The answer is no.
In fact, manually labeled data is less prone to errors regarding quality assurance and data exceptions.
The human workforce cannot be totally replaced by some tools leading with an AI-based automation feature, especially dealing with exception, edge cases, complex data labeling scenarios, etc.
ByteBridge is a human-powered and ML-powered data labeling tooling platform. We provide scalable, high-quality training data for ML/AI industry with flexible workflow.
Quality and Accuracy
- ML-assisted capacity can help reduce human errors by automatically pre-labeling
- The real-time QA and QC are integrated into the labeling workflow as the consensus mechanism is introduced to ensure accuracy.
- Consensus — Assign the same task to several workers, and the correct answer is the one that comes back from the majority output.
- All work results are completely screened and inspected by the machine and human workforce.
In this way, ByteBridge can affirm our data acceptance and accuracy rate is over 98%.
- Progress preview: client can monitor the labeling progress in real-time on the dashboard
- Result preview: client can get the results in real-time on the dashboard
- Real-time Outputs: client can get real-time output results through API. We support JSON, XML, CSV, etc., and we can provide customizable datatype to meet your needs.
3D Point Cloud Annotation Service
ByteBridge self-developed 3D Point Cloud labeling, quality inspection tool, and pre-labeling functions can complete high-quality and high-precision 3D point cloud annotation for 2D-3D fusion or 3D images provided by different manufacturers and equipment, and provide one-station management service of labeling, QA, and QC.
More info: ByteBridge Launches World’s First Mobile 3D Point Cloud Data Labeling Service
3D Point Cloud Annotation Types:
- Sensor Fusion Cuboids: 12 categories include car, truck, heavy vehicle, two-wheeled vehicle, pedestrian, etc.
- Sensor Fusion Segmentation: obstacles classification, different types of lanes differentiation
- Sensor Fusion Cuboids Tracking
① Tracking the same object with the same ID, labeling the leaving state;
② Point clouds or time-aligned images could be provided, point clouds outputs only.
A collaboration of the human-work force and AI algorithms ensure a 50% lower price compared to the conventional market.
As the quality of the labeled dataset determines the success of the self-driving industry, cooperation with a reliable partner can help developers to overcome the data labeling challenges.
We can provide personalized annotation tools and services according to customer requirements.
If you need data labeling and collection services, please have a look at bytebridge.io, the clear pricing is available.
Please feel free to contact us: email@example.com
1 Introduction of Giant Self-driving Companies: Tesla and BAIC
2 Didi Autopilot Has Completed a New Round of Financing of over US $300 million
3 3D Point Cloud Annotation is Indispensable in Boosting Self-Driving Industry?
4 Will Cruise’s Commercial Test be Far Behind When it Realizes the Full Driverless Vehicle Carrying Passengers?
5 Eight Common Data Annotation and Labeling Tools in Autonomous Vehicle Industry
Source: financial 11