Robotic Vision Scene Understanding Challenge



The Robotic Vision Scene Understanding Challenge evaluates how well a robotic vision system can understand the semantic and geometric aspects of its environment. The challenge consists of two distinct tasks: Object-based Semantic SLAM, and Scene Change Detection.

Key features of this challenge include:

Watch the video below to learn more!

Challenge Tasks

The semantic scene understanding challenge is split into 3 modes, across 2 different semantic scene understanding tasks, for a total of 6 different challenge variations. The different variations open up the challenge to a wider range of participants, from those who want to focus solely on visual scene understanding to those wish to integrate robot localisation with visual detection algorithms.

The two different semantic scene understanding tasks are:

  1. Semantic SLAM: Participants use a robot to traverse around the environment, building up an object-based semantic map from the robot’s RGBD sensor observations and odomtry measurements.
  2. Scene change detection (SCD): Participants use a robot to traverse through an environment scene, building up a semantic understanding of the scene. Then the robot is moved to a new start position in the same environment, but with different conditions. Along with a possible change from day to night, the new scene has a number objects added and / or removed. Participants must produce an object-based semantic map describing the changes between the two scenes.

The object-based semantic maps generated by submissions are evaluated against the corresponding ground-truth object-based semantic map (for task 2 this is the map of changes in the second scene, with respect to the first). Please see the BenchBot Evaluation documentation for details on object-based semantic maps submission formats, and further explanation of the two different tasks.

Each task has 3 variations, corresponding to the following modes:

  1. Passive control, with ground-truth localisation: The robot follows a fixed-trajectory, and participants are given a single method to control the robot: moving to the next pose. The task cannot be continued once the entire trajectory has been traversed. Participants receive ground-truth poses for all robot components after each action.
  2. Active control, with ground-truth localisation: The robot can be controlled by either moving forward a requested distance, or rotating on the spot a requested number of degrees. The task cannot be continued if the robot collides with the environment. Participants receive ground-truth poses for all robot components after each action.
  3. Active control, using dead reckoning (no localisation): The robot can be controlled by either moving forward a requested distance, or rotating on the spot a requested number of degrees. Participants receive poses derived from robot odometry after each action, with localisation error that accumulates over time.

Please see the BenchBot API documentation for full details about what actions and observations are available in each mode, and how to use them with your semantic scene understanding algorithms.

How to Participate

Our challenge is currently live (since June 2020), and available here on the EvalAI website. Please create an account, sign in, & click the “Participate” tab to enter our challenge. Full details on how to participate, the available software framework, and submission requirements are provided on the site.

The challenge is open until the 2nd of October 2020, and a total of $2,500 USD will be awarded to high-performing competitors. The two best performing teams also receive a Titan RTX GPU and up to 5 Jetson Nanos (1 per team member) provided by our sponsors Nvidia.

Participating in the Semantic Scene Understanding Challenge is as simple as the 4 steps below. The BenchBot software stack is designed from the ground up to eliminate as many obstacles as possible, so you can focus on what matters: solving semantic scene understanding problems. A collection of resources, documentation, and examples are also available within the BenchBot ecosystem to support your experience while participating in the challenge.

To participate in our challenge:

  1. Download & install the BenchBot software stack. Use the examples to dive straight in & start playing.
  2. Choose a task to start working on a solution for, using benchbot_run --list- tasks to list supported tasks
  3. Start with the development environments “miniroom” and “house” which include ground-truth maps to aid in your algorithm development
  4. Develop a solution using the benchbot_run, benchbot_submit, & benchbot_eval scripts
  5. Create some results for your solution in the challenge environments using: benchbot_batch -t <your_task> -E <benchbot_root>/batches/challenge/<your_task> -z -n <your_submission_cmd>
  6. Use the Submit tab at the top of our EvalAI page to submit your results for evaluation


Talk to us on Slack or contact us via email at

Organisers, Support, and Acknowledgements

Stay in touch and follow us on Twitter for news and announcements: @robVisChallenge.

David Hall
Queensland University of Technology
Ben Talbot
Queensland University of Technology
Haoyang Zhang
Queensland University of Technology
Suman Bista
Queensland University of Technology
Rohan Smith
Queensland University of Technology
Feras Dayoub
Queensland University of Technology
Niko Sünderhauf
Queensland University of Technology

The Robotic Vision Challenges organisers are with the Australian Centre for Robotic Vision and the Queensland University of Technology (QUT) Centre for Robotics in Brisbane, Australia.