The Track
Symmetry is a common characteristic found in many aspects of the world. It can be observed in natural objects such as animals, plants, and even humans, as well as in man-made objects. The prevalence of symmetry in artificial objects can often be attributed to its perceived connection to beauty. In addition, symmetry has had a significant impact on fields such as art and engineering. Therefore, it is essential to understand the concept of symmetry from various perspectives in order to effectively utilize this understanding to solve problems.
Symmetry is a useful concept in reducing complexity, as it allows us to represent an entity with less information. This makes the analysis of symmetries appealing in the search for more compact representations of objects in computer systems. However, computational representations do not inherently contain information about their possible symmetries, so it is necessary to analyze these representations and search for symmetries using the available information. The task of detecting symmetries computationally is challenging and has received significant attention from various research communities, such as those in computer vision and geometric processing.
The purpose of this track is to assess the effectiveness of automatic algorithms in detecting symmetries in 3D point clouds that represent simple shapes. The objective is to determine the reflective planes for each point cloud. By simple shapes, we refer to surfaces generated by various types of closed plane curves that are projected or rotated in various ways. Specifically, a closed plane curve can be utilized as the directrix of a cylinder or a cone, or it can be rotated relative to an axis.
The dataset
A dataset consisting of 3D simple shapes represented as point clouds will be provided, divided into a training set and a test set. For each point cloud in the training set, we will provide the normal vectors to the reflective planes and a point that lies on the reflective plane. Additionally, we will apply various transformations (such as noise and undersampling) to the point clouds to evaluate the robustness of methods for detecting symmetries under these transformations. Examples of 3D point clouds and their reflective symmetries can be seen in the following figure.
Downloads
The data can be downloaded in this link. The training set contains the point clouds and the ground-truth files. Each ground-truth file contains the number of symmetries for a given point cloud. For the symmetries, each line contains three float numbers for the normal and three float numbers for the point-in-plane. The test dataset contains only the point clouds.
Results submission
Participants should submit a set of text files with the results of their runs. For a given test file, participants must submit a text file with the detected symmetries. The format of the submitted file is as follows:
- The first line of the text file must contain the number of detected symmetries for the test point cloud.
- There must be one line per detected symmetry. Each line contains seven numbers. The first three numbers are the normal of the detected plane. The last three numbers are the point belonging to the detected plane. The last number is the confidence of the detected symmetry. The confidence must be a number in the range [0,1]. The confidence will be used to rank the detections and to compute the evaluation metrics. If your method does not compute confidences, you can put a value of 1.0.
In addition, participants must report the following information:
- System specification: CPU (model, speed in GHz, number of CPU's, RAM per CPU in GB). In case participants use GPU, the required information is model, speed in MHz, memory in GB and number of GPU's.
- Processing time in seconds: In order to properly evaluate the several stages of the method, participants should make the difference between offline processing (for example neural network training) and online processing (the time for inference). Please, provide the average query time for the Nq query objects.
Effectiveness evaluation
The performance of the algorithms will be evaluated on the basis of the quality of the recognised planes and the number of planes correctly identified. This analysis will be specified according to the type of point cloud transformations, in order to provide a thorough evaluation of the robustness of the algorithms. The metric to compare the methods will be the mean average precision (mAP) of the detections, similar en spirit to the evaluation of object detection methods in computer vision.
To Participants
People interested in participating in this track must register by sending an email to Ivan Sipiran (isipiran@dcc.uchile.cl) and Chiara Romanengo (chiara.romanengo@ge.imati.cnr.it). The registration will help to keep track of the contest.
Timeline
- January 6th, 2023: Track kick-off
- January 18th, 2023: Deadline for track registration
- January 18th, 2023: The dataset will be available for downloading
- February 28th, 2023: Deadline for submission of results
- March 28th, 2023: Submission of track paper
Contact
For additional information, please do not hesitate to contact Ivan Sipiran (isipiran@dcc.uchile.cl).
Organizers
- Ivan Sipiran, Department of Computer Science, University of Chile.
- Chiara Romanengo, Istituto di Matematica Applicata e Tecnologie Informatiche "E. Magenes", Consiglio Nazionale delle Ricerche, Italy.
- Silvia Biasotti, Istituto di Matematica Applicata e Tecnologie Informatiche "E. Magenes", Consiglio Nazionale delle Ricerche, Italy.
- Bianca Falcidieno, Istituto di Matematica Applicata e Tecnologie Informatiche "E. Magenes", Consiglio Nazionale delle Ricerche, Italy.