Smartathon

16788 Registered Allowed team size: 1 - 6
16788 Registered Allowed team size: 1 - 6

This campaign is over.

Prototype Phase
Online
starts on:
Jan 08, 2023, 03:00 PM UTC (UTC)
ends on:
Jan 21, 2023, 09:00 PM UTC (UTC)

Overview

About SDAIA
SDAIA’s transformation strategy was approved in 2019. The strategy gives
SDAIA a core mandate to drive and own the national data and AI agenda to
help achieve Vision 2030’s goals and our Kingdom’s highest potential. To
fulfill this mandate, SDAIA – and its sub-entities NDMO, NIC, and NCAI – will
deliver on the promise to create a data-driven and AI-supported government
and economy.

Mission
Unlock the value of data as a national asset to realize Vision 2030
aspirations by setting the national data and AI strategy and overseeing its
execution through harmonized data policies, data analytics and insights
capabilities, and continuous data and AI innovations.

Vision
Positioning the Kingdom as a global leader in the elite league of data-driven
economies.

About Smartathon
A solution-based challenge launched by SDAIA together with several partners, it aims to develop innovative solutions and detect visual pollution in the cities of the Kingdom. Smartathon aims to enhance the deduction of visual pollution and improve the accuracy of existing algorithm.

Themes

Please go through the SUBMISSION GUIDELINE tab for both the themes.


If you selected the "Theme 1" or "Both" option during registration, your Team Leader will get an invite for the ML platform to submit your solution. Please check your email.

 

For Theme #1:

  • You may submit your project as many times as you like. Only the final submission will be judged.
  • Make your submission on the ML platform in .csv format.
  • Your model will be evaluated along with your final submission. SDAIA reserves the right to reject any submission if the model is found to be sub par..
  • All projects must contain the following in their submission as a simple writeup of maximum 800 words:
    • Project overview (3-4 sentences explaining what you built)
    • Describe what challenges you faced at preparing the data and how you approached the problem.
    • Describe how you would make your solution more scalable.
    • Declare what open source software you used to build your solution.
    • Describe what you would want to try if you had more time / resources / data to resolve the problem.

 

For Theme #2:

 

  • All projects must contain the following in their submission as a simple writeup of maximum 1000 words:
  1. Describe how you build a system that is able to effectively localize the potholes?
  2. Describe how you are able to estimate the magnitude (in depth and area) of the potholes detected?
  3. Tell us if you managed to represent the pothole characteristics via some sort of 3d reconstruction from the visual data alone?
  4. Describe how your proposal shows potential in:
    1. Accurately detecting potholes
    2. Accurately estimating the magnitude of the potholes
    3. Accuracy establishing the most urgent or problematic potholes in the analyzed data.
    4. Potential maturity in the solution as an alternative to Mobile Lidar on pothole detection via Computer Vision.
    5. Level of solution autonomy. How much human intervention does the analysis need? Does it potentially scale well?
Detection and evaluation of the following elements on street imagery taken from a moving vehicle

Since visual pollution is a relatively recent issue compared to other forms of environmental
contamination, study is needed to define, formalize, measure, and evaluate it from many angles.
This competition aims to establish a new field of automated visual pollution classification,
utilizing the technological prowess of the twenty-first century for environmental management
applications.
By training and testing approaches to convolutional neural networks we expect competitors to
simulate the human learning experience in the context of picture identification for the
classification of visual pollutants.
Additionally this will be useful for the development of a "visual pollution score/index" for urban
areas that might produce a new "metric" or "indicator" in the discipline of urban environmental
management.
In this competition, you will build and optimize algorithms based on a large-scale dataset. This
dataset features the raw sensor camera inputs as perceived by a fleet of multiple vehicles in a
restricted geographic area in KSA
If successful, you’ll make a significant contribution towards stimulating further development city
planning and empowering communities around the world.


Visual pollution types:


● GRAFFITI
● FADED SIGNAGE
● POTHOLES
● GARBAGE
● CONSTRUCTION ROAD
● BROKEN_SIGNAGE
● BAD STREETLIGHT
● BAD BILLBOARD
● SAND ON ROAD
● CLUTTER_SIDEWALK
● UNKEPT_FACADE

Dataset: https://drive.google.com/file/d/1ULqYtd9yomeGz53WBhgRdPRFB37ppeDU/view

*If you selected for "theme 1" or "Both" option for theme during registration process, your Team Leader will get the invite for "Theme 1" ML platform where you can submit your solution.

Pothole severity classification via computer vision

The majority of current pavement condition assessment techniques are labor-intensive and
manual. Existing techniques for identifying and evaluating potholes rely on 3D surface
reconstruction, which is expensive in terms of both hardware and compute, or on acceleration
data, which only yields preliminary results. We are looking for teams to propose a low-cost
method for automatically identifying potholes and judging their severity using vision-based data
for both 2D and 3D reconstruction.


Utilizing the visual and spatial qualities of potholes as well as the measurement parameters
(width, quantity, and depth) that are used to estimate pothole severity, we would like to
understand how both 2D and 3D reconstruction can be combined to improve recognition results.
While the width and depth of the potholes are determined using 3D reconstruction, the number
and location of potholes is determined using 2D recognition.


Data gathering, distress identification and classification, and distress evaluation are the three
steps in the pavement assessment process. Modern data collection techniques are being quickly
replaced by inspection vehicles. These inspection vehicles can gather data at speeds of up to 60
mph (96 km/h), thanks to their numerous sensors, which include cameras for surface imaging,
optical sensors for distance measurement, laser scanners for profiling, ultrasonic sensors for
rutting detection, and accelerometers for roughness measurements. The second and third
processes of distress classification and assessment are still mostly manual despite the
automation of the data collection procedure. Currently, technicians manually examine the
collected data to determine the existence of distresses and gauge their severity from the
computer screen. Such a labor-intensive manual method might become unsystematic due to the
volume of data that needs to be collected, which eventually lowers the assessment's quality.
Although there are clear rules for manual diagnosis and assessment of asphalt distress, the
technicians' experience affects the assessment's outcome. A hybrid imaging device that
combines digital cameras and infrared lasers to capture continuous images of lines projected by
infrared lasers is based on 3D surface profiles from time-of-flight laser scanners to classify and
quantify pavement deterioration. These commercial software programs do not, however, count
or identify all the potholes that have been spotted.

Dataset:

https://drive.google.com/file/d/1iYgiw4B4uHG78trDbJ9n63ZKCEH96vvx/view

Challenge:

We challenge the community to explore (1) novel approaches using computer vision for
classification of pothole severity in roads and (2) propose potential pipelines that would allow
city planners to assess pothole severity in roads in a cost-effective manner.
Your challenge is to generate actionable, practical, and novel insights from video and picture
data of roads that devise innovative and data-driven approaches to analyzing road conditions,
prioritization of work to be done on roads, and classification of severity of potholes and cracks
on roads.

There are several potential items to analyze:


These include, but are not limited to:
● Pipeline that automatically detects the bounding boxes or segments of the potholes
within the visual footage.
● Pipeline that automatically creates a 3d point cloud reconstruction of the road, including
the potholes.
● Use of both 2D and 3D data to automatically classify severity and characteristics of the
potholes detected, and based on this data determine which potholes are top priority for
resolution.
● Accuracy of the measurement in terms of:
○ Width and Height of the Pothole (in cms)
○ Depth of Pothole (in cms)
○ Estimated shape of the pothole detection.
● Metric that determines the segments of the road that require the most immediate
attention.

A potential approach

PLEASE NOTE: Contestants should not feel limited to these suggestions.


First, videos captured by a High Definition camera are utilized to look for potholes.
Simultaneously, a sparse first 3D reconstruction is performed using the same footage. Potholes
are confirmed to exist in order to decrease the amount of places that are incorrectly classified as
potholes based on the findings of 2D detection and 3D sparse reconstruction. Next, a dense
reconstruction approach is used to enhance the results of the sparse 3D reconstruction. The
geometrical characteristics of the potholes and their severity are measured using the dense 3D
point cloud model and the output from the 2D appearance-based recognition.

PLEASE NOTE: Contestants should not feel limited to these suggestions.

Based on the 2D localization data the location of the Pothole within the frame can be determined
using YOLO, Faster R-CNN, or any other Computer Vision based object localization method.
Teams are free to propose their preferred method.
3D reconstruction of the scene could be done using Computer Vision techniques. Teams may
propose any algorithm to generate 3d point clouds and later meshes from overlapping pictures
of the road. Any available structure from motion algorithms or software can be used, including
commercially available packages and packages available in the open source. Most commercial
alternatives offer free trials.

Scoring
An entry to the competition consists of a Notebook submission that is evaluated on the following
five components, where 0 is the low score and 10 is the high score. Submissions will be judged
based on how well they address:
Innovation:
● Are the proposed approaches actionable?
● Is this a way of evaluating road conditions that is novel?
● Is this project creative?
Accuracy:
● Are the results similar to ground truth data?
● Is the method by which these results obtained clearly explained?

Relevance:
● Would Cities benefit from using this approach?
● Would this potentially be able to scale in a cost-effective manner versus standard
methods of road condition evaluation?

Judges will consist of city planners, and computer vision specialists that are working for either
Saudi Data and Artificial Intelligence Authority (SDAIA), Ministry of Municipal, Rural Affairs, and

Housing, and Royal Commission for Riyadh City. Scores will be averaged so that each of the
components above is weighed equally.

Notebook requirements

All notebooks submitted must be made public on or before the submission deadline to be
eligible. If submitting as a team, all team members must be listed as collaborators on all
notebooks submitted.

 

Prizes SAR 1,000,000 in prizes

Main Prizes
1st Prize
SAR 375,000

Theme 1

2nd Prize
SAR 112,500

Theme 1

3rd Prize
SAR 37,500

Theme 1

more

Social Share

Help & Support

Please contact event admin
HackerEarth Support at support@hackerearth.com
Notifications
View All Notifications

?