‘Conductor ka khoon, driver ka paseena…road par chalti hai ban kar haseena.’
Such messages painted on Indian trucks can be funny but also at times very close to the truth. Indian roads are chaotic, dangerous and tough for humans to master. For algorithms that are making some of the autonomous driving possible, it’s even harder simply because there isn’t enough data on Indian road conditions to learn from.
But that’s about to change. The world’s first public dataset (and the biggest of its kind) of Indian driving conditions has been readied by a computer vision scientist and his team at the IIIT-Hyderabad, one among the country’s two dozen International Institutes of Information Technology. The dataset, which is now being made public, will help computer vision researchers from across the world to train their algorithms on Indian driving conditions, eventually leading to solutions that could make roads safer.
“Modern AI is dominated by the use of data and if you want Indian problems to get attention worldwide, we need data and this is something that will be the fundamental building block,” says C V Jawahar, who heads the computer vision group at IIIT Hyderabad. The idea is to make this data openly accessible for researchers to train their machine learning models.
The 35-member team has been mostly using images and video captured using simple cameras attached to vehicles. Though there are sensors out there that can capture similar data, using cameras keeps the costs low.
“People would like to eventually move to video or image-based navigation as they are cheaper. We want solutions that are cheap and affordable to be able to democratise this data,” says Jawahar, who focuses his energy mostly on solving problems that are relevant to Indian situations.
To be sure, there has been data collected in the past but a large part of that is not publicly available. Moreover, data collected by companies or startups often tend to be incompatible with other datasets for it to be of any use to the larger community.
“This is the holy grail of data sets and will test the best in class algorithms,” says Dheemanth Nagaraj, an Intel fellow and architect, server CPU development and new products innovations at chipmaker Intel. The project, funded by Intel, began in November 2017 in partnership with the government of Telangana and Karnataka.
The team has been driving around in these states, collecting drive sequences, and then annotating them using a computer vision algorithm. The idea is to label signposts, pedestrians, types of vehicles, street lights and so on.
“The thought process here is to enhance road safety,” says Nagaraj. “But the first step is data sets. The goodness of our AI systems is based on the potency of the data sets.” How does this tie into road safety? You could train a computer vision model to conduct infrastructure audits — check if the street lights are on or dividers are in place — using this dataset.
The idea was to create a dataset that’s double the size of Cityscape, the biggest publicly available dataset at the time. Using over 180 drive sequences, the team created about 10,000 pixel-level annotated images and 50,000 object level annotated images. Cityscape has a dataset of about 5,000 frames annotated at the pixel level for 50 cities in Germany. Pixel-level annotation means each pixel in the image is associated to an object class such as road, rider, guardrail, car, truck, bus, sky, and so on. Since then, UC Berkeley, Baidu and a few others have released larger datasets mostly focussed on US and China.
The team from IIIT Hyderabad plans to run a challenge at the upcoming European Conference on Computer Vision (ECCV) to attract more researchers to use the dataset. “Now the dataset is being made public. This is useful for non-commercial use and research,” says Jawahar. The Indian Driving Dataset (IDD), which has about 26 classes, will put Indian road safety problems on the radar of AI researchers.
“Our roads are an order harder than the western roads and algorithms trained on western datasets won’t perform well in our conditions,” says Jawahar. The IDD will have additional classes to identify auto rickshaws, odd-shaped vehicles, riders and pillions on a two-wheeler etc. “This itself will increase the accuracy of algorithms,” adds the professor who plans to continue the project for a “few more years” so that other aspects related to road safety such as overtaking and wrong side driving can be captured.
Will India be ready for autonomous driving then? “There’s a bit of hype about autonomous driving. It works in limited settings but I don’t think even roads in the US are not going to be dominated by autonomous cars in the next five years or so,” says Jawahar.
In a major setback to autonomous testing, the state of Arizona banned Uber’s self-driving cars in March after one such car killed a pedestrian in a crash. Massachusetts and a few other US cities have also asked companies to take autonomous cars off the streets. The crash has also forced companies to take what The Economist calls a “more realistic” route to autonomous driving. Drive.ai, for instance, which operates autonomous minivans in Frisco in Dallas, tries to differentiate self-driving cars by design (painting them Orange) or operating in daylight hours. The idea is to take small steps before aiming for level-5 autonomy. Car automation is classified into five levels— ranging from 0 which is for cars that have no automation to level 5 which is full automation or the equivalent to human drivers.
In an Indian context where there is no shortage of manpower, more than making cars autonomous, priority must be given to make roads safer, insists Jawahar. “Autonomy will come step by step. We’ll see semi-automatic systems, driving assistants, interactive systems and safety features that are AI enabled before that,” he says.
Making such datasets has big advantages. “It’s difficult to build a model without good data. There are similar efforts in areas such as biology and healthcare going on across the world,” says Thejesh GN, an independent technologist and the founder of DataMeet, a community of data scientists and open data enthusiasts. iNaturalist, for instance, is a dataset that identifies species; the COCO dataset is a large-scale object detection, segmentation, and captioning dataset; and the Tumor Proliferation Assessment Challenge 2016 dataset helps build models for cancer detection. “For a startup to collect such data, annotate them..label them and categorise them… it takes a while. So this is really really useful,” says Thejesh.
Subscribe to FactorDaily
Our daily brief keeps thousands of readers ahead of the curve. More signals, less noise.
Updated at 08:21 am on September 11, 2018 to change the headline for better syntax. The earlier headline read: "World's first open, public traffic dataset of Indian roads being readied in crowded Hyderabad".
Disclosure: FactorDaily is owned by SourceCode Media, which counts Accel Partners, Blume Ventures, Vijay Shekhar Sharma, Jay Vijayan and Girish Mathrubootham among its investors. Accel Partners and Blume Ventures are venture capital firms with investments in several companies. Vijay Shekhar Sharma is the founder of Paytm. Jay Vijayan and Girish Mathrubootham are entrepreneurs and angel investors. None of FactorDaily’s investors has any influence on its reporting about India’s technology and startup ecosystem.