Your address will show here +12 34 56 78
AI

Autonomy is quickly becoming the driving force of the automobile industry. Leading autonomous car makers – GM, Waymo, Tesla, Ford, etc. – are all aggressively improving the infrastructure necessary for autonomy. As there are multiple autonomous vehicle technologies at play, a hot discussion is going on in the market about which one is better – vision based (camera) or sensors based (LiDAR, radar, and ultrasonic).

Currently, pretty much all active players in the market, with the exception of Tesla, are moving forth with LiDAR. Tesla, on the other hand, is betting solely on cameras/vision.

Before we get into thick of the discussion, let’s quickly go through these automotive vehicle technologies (LiDAR, radar, and camera).

LiDAR

LiDAR (Light Detection and Ranging) devices are basically active sensors that emit high-frequency laser signals in quick succession (up to 150k pulses/sec). It measures the time taken for each signal to bounce back and calculate the distance between the vehicle and obstacles with high precision. The main concerns with LiDAR are that it is way too costly and doesn’t recognize colors and traffic signs.

future of autonomous vehicle technology LiDAR

Radar

LiDAR and Radar are essentially the same technologies operating with light waves at different frequencies. Radar emits low-frequency signals; therefore, produce less accurate outcomes (see the image below). While cost is not an issue with radar technology, being a sensor, it also doesn’t recognize colors and signs, which are crucial for autonomous driving.

future of autonomous vehicle technology lidar vs radar

Ultrasonic is another automotive vehicle technology in wide use today. Unlike LiDAR and radar, these sensors use soundwaves to measure the distance from obstacles.

Camera

Being a vision-based device, camera estimates the distance based on its relative size in the captured frame. The main advantage with camera is that it can recognize traffic signs and traffic light colors, which sensors cannot. While cost is another advantage with camera, but given the amount of efforts the goes into training camera-based systems to achieve sensor level accuracy, it is not a very cost-effective solution either.

future of autonomous vehicle technology computer vision

Related read: Computer Vision Trends 2019

Now that we have brushed up the basics of main autonomous vehicle technologies, let’s see how they are faring in the real world.

Current Industry State

It is clear that one technology – be it vision or sensors – cannot provide the ultimate solution for driverless cars. As a senior analyst at Navigant puts in “Layering sensors with different capabilities, rather than just relying on a purely vision-based system, is ultimately a safer and more robust solution”.

So, the question here is, which one should be used as the primary autonomous vehicle technology, and which one as the secondary? As mentioned earlier, LiDAR is the industry’s first choice for driverless cars. Main reason being, cameras estimate the distance between objects based on relative size; however, LiDAR sensors know it with high precision.

However, Tesla’s CEO Elon Musk, who has long criticized LiDAR, recently called it “a fool’s errand” at Tesla’s Autonomy Day and claimed that everyone will drop it in the near future. As absurd as it sounds, Musk backed his claim with some reasoning (although it may sound unreasonable to most). First, LiDAR is costly hardware. Secondly, he demonstrated how camera is getting better and better at achieving sensor-level accuracy.

future of autonomous vehicle technology vision vs sensors

Currently, cost is undeniably the biggest let-down with LiDAR; however, all active players are putting an enormous amount of efforts to bring its cost down. For instance, in 2017, Google-backed Waymo claimed that it will build its own LiDAR sensors and will reduce the cost from $75,000 to $7,500 apiece. In March 2019, Waymo updated on its claim that it will soon start selling its in-house LiDAR devices to third-parties.

Another drawback, LiDAR becomes inaccurate in fog, rain, and snow, as it’s high-frequency signals detects these small particles and include them in the rendering. On the other hand, radar works quite effectively in such weather conditions and it also costs considerably less; therefore, is arguably a better substitute than LiDAR when used as secondary autonomous vehicle technology in conjunction with cameras.

Also Read: The Current State of AI in the Corporate World

What Future Holds for Autonomous Vehicle Technology

Given the ongoing developments, it is likely that the future will unfold either of the following two possibilities:

  1. Camera-based vision systems achieve LiDAR level accuracy
  2. LiDAR becomes cheaper than the cost of improving camera systems

Which will happen first is hard to say with any certainty. LiDAR prices have dropped over the last few years, so has the camera’s image processing. Tesla, with its Fleet Learning capabilities, is betting high on cameras and argues that its systems are designed to replicate human behavior and recognize the world around with colors and signs, not by distance. However, discarding LiDAR completely is also not an option at the current stage. So, the next few years will be crucial in deciding where the course of autonomous vehicles turns.

0

AI

Why are machines so smart?

Because we make them so.

But they can be only as smart as we make them to be.

In today’s AI ecosystem, there is pressure on everyone to make their machine learning algorithms as good as human intelligence. And the only way to achieve it is to have a good amount of quality labeled data to train those algorithms.

Except that the data doesn’t come easy.

Every organization entering into machine learning related services faces this challenge today. And to overcome it, they must have the know-how of different data labeling tools that can help in building quality training data sets and build them efficiently.

Here is a list of the best data labeling tools based on what type of data you are labeling.

Image & Video Labeling Tools

Annotorious

It’s a free, easy to use, MIT-licensed annotation tool for labeling of images on a website. It’s free for commercial use as well. Integrating it with your website only requires adding 2-3 lines of code. You can also explore many of its features in the demos.

annotorious image video data labeling tool 2019

Annotorious’ most noticeable features include:

  • Image annotation with bounding boxes
  • Process maps and high-resolution zoomable images
  • Annotorious can be modified with plugins to suit a particular project’s need
  • Annotorious Community; where developers can find how they can modify it to extend its capabilities
  • Annotorious Selector Pack plugin (to be launched), which will include features like custom shape labels, freehand, point, and Fancy box selection

LabelMe

LabelMe is an open source online data labeling tool. With simple signup, it allows users to label images and share their annotation publically, which is primarily used for a range of computer vision based applications and research.

LabelMe image video data labeling tool 2019

Some of LabelMe’s key features include:

  • LabelMe also offers its mobile app for image labeling & annotation
  • Image collection, storage, and labeling
  • Training object detectors in real-time
  • Simple and intuitive UI
  • LabelMe offers MATLAB Toolbox that allows users to download and interact with the images and annotations in the LabelMe database

Labelbox

Labelbox is one of the most versatile labeling tools available today. Its comprehensive features enable organizations to easily adapt and train their machine learning models. Its pricing varies based on the amount of data and the sophistication of the model you are training.

labelbox image video data labeling tool 2019

Key features include:

  • Labelbox supports Polygon, Rectangle, Line, and Point segmentation, as well as pixel-wise annotation
  • You can create bounding boxes and polygons directly on the tiled imagery (zoomable maps)
  • Ideal to work with a big team of labelers as it serves up images to be labeled asynchronously, i.e., no two labelers label the same image
  • Assured security as the source data is either stored in-house or on a private cloud
  • Labelbox allows you to maintain quality standards by keeping track of labeling task performance

Sloth

Sloth is a versatile annotation tool for various data labeling tasks related to computer vision research. It’s free and is one of the most popular tools for facial recognition, therefore, is widely used for surveillance and user identification related applications.

sloth image data labeling tool 2019 facial recognition

Most notable of Sloth’s features include:

  • It allows an unlimited number of labels per image or video frame – leading to more detailed file processing
  • It supports various image selection tools – points, rectangles, and polygons
  • Developers consider Sloth as a framework and set of standard components that can be configured to build a label tool specifically tailored to one’s needs

Audio Labeling Tools

Praat

Praat is a free audio labeling tool under the Creative Commons (CC BY SA) license, meaning, any derivative works must also come under creative commons license.

praat audio data labeling tool 2019

Praat’s primary features include:

  • Spectral analysis, pitch analysis, format analysis, and intensity analysis of audio files
  • It can also identify jitter, shimmer, voice breaks, cochleagram, and excitation pattern
  • You can work with sound files of up to 3 hours (2GB)
  • It allows you to mark time points in the audio file and annotate these events with text labels in a lightweight and portable TextGrid file
  • Users can work with sound and text files at the same time when text annotations are linked with an audio file

Aubio

Aubio is another free and open source annotation tool for audio data labeling. The tool is designed to extract annotations from audio signals. Aubio is written in C and is known to run on most modern architectures and platforms.

aubio audio data labeling tool 2019

Aubio offers the following key features:

  • Digital filters, phase vocoder, onset detection, pitch tracking, beat and tempo tracking, mel frequency cepstrum coefficients (MFCC), transient / steady-state separation
  • You can segment a sound file before each of its attacks, performing pitch detection, tapping the beat and producing midi streams from live audio
  • There’s a dedicated function library to execute above-mentioned functions in real-time applications
  • Users can also use these functions offline via sound editors or software samplers

Speechalyzar

Speechalyzar is an audio data labeling tool specifically designed for the daily work of a ‘speech worker’. It can process large speech data sets with respect to transcription, labeling, and annotation. Its main application is the processing of training data for speech recognition and classification models.

speechalyzar audio data labeling tool 2019

Speechalyzar’s main features include:

  • You can implement it as a client-server based framework in Java and interfaces software for speech recognition, synthesis, speech classification and quality evaluation
  • Speechalyzar also allows you to perform benchmarking tests on speech-to-text, text-to-speech and speech classification software systems
  • Ideal for manual processing of large speech datasets

EchoML by Azure and Soundscape are some other audio data labeling tools with rich visualization capabilities that you can also explore.

Suggested: Leverage the power of data with visualization

Text Labeling Tools

Rasa NLU

Rasa NLU is an open-source NLP tool for intent classification and entity extraction. It is primarily used to annotate text for chatbots but can be used for a variety of applications. For instance, recently Trantor used Rasa NLU to train a machine learning model to detect harassment and abuse in email communication within an organization.

Rasa NLU text data labeling tool 2019

Some of the advantages with Rasa NLU are:

  • Users can tag multiple words in a single sentence to their respected class or assign the same word in multiple classes
  • You can customize and train its language model as per domain-specific needs and get higher accuracy
  • Rasa NLU’s open source library runs on premise to keep users’ data safe and secure

Stanford CoreNLP

Stanford CoreNLP is a free, integrated NLP toolkit that provides a set of human language technology tools, which allow users to accomplish various text data pre-processing and analysis tasks.

Stanford CoreNLP text data labeling tool 2019

Here are some advantages with Stanford CoreNLP:

  • It offers a broad range of grammatical analysis tools (base forms of words, parts of speech, names, normalize dates, times, and numeric quantities, mark up the structure of sentences in terms of phrases and syntactic dependencies)
  • CoreNLP is a fast, robust annotator for arbitrary texts, widely used in production
  • It offers a modern, regularly updated package, with the overall highest quality text analytics
  • Support for a number of major (human) languages
  • It offers APIs for most major modern programming languages and can run as a simple web service

Bella

Bella is a text annotation tool that helps data scientists manage, label, and evaluate natural language datasets. It is designed to save time spent in measuring and learning data, which involves collecting, inspecting, training, and testing data.

Bella text data labeling tool 2019
Image: A Bella project file for labeling a social media post (Source: Github)

Some plus points of using Bella:

  • It offers an intuitive GUI, which allows users to label and tag data through convenient keyboard shortcuts and swipe, and visualize metrics, confusion matrices, and more
  • Bella also offers database backend to easily manage labeled data
  • Bella is a preferred tool for sentiment analysis, text categorization, entity linking and POS tagging

Tagtog

Tagtog is a versatile text labeling tool that offers manual as well as automated annotation. It’s an AI startup with an impressive client base including AWS, Siemens, and a number of data science research institutions.

tagtog text data labeling tool 2019

Some of the best things about Tagtog are:

  • Users don’t require coding knowledge or data engineering concepts to use Tagtog
  • Tagtog offers inbuilt ML model to automate text annotation and also provides hassle free deployment and maintenance of manually trained models
  • Tagtog annotation tool allows multiple users to collaborate to a single project

Conclusion

There are numerous other data labeling tools in the market, apart from the ones listed above. And as with any other tool for any other purpose, the key is not to know a lot of tools but to know which tool will work best for a given project and to understand how to leverage it best.

And as for which approach you should adopt for labeling – in-house or outsource – that also depends on the project requirements. If you have time and resources, you can do it in-house. If priority is to cater customers with AI driven solutions as quickly as possible, then it is suggested to outsource your projects to a professional firm.

The machine learning market is brewing up and companies are in a rush to get ahead of each other. So, in current dynamics, spending a little to take the advantage of the early bird can make a big difference in the long run.

0

Talk to Our Solution Expert

Contact Us