2026 Munich Tech Arena

 

Access to Example Code, Sample Data, and Proposal Template

 

The goal of this challenge is to extract the landmarks for the pinna shape from the 3D scans of a human subject. The submitted models are expected to accept the 3D mesh as the input and output precise landmarks of the pinna shape. These landmarks represent the outer helix, concha outline, inner helix, and superior antihelix [1] as illustrated in Figure 1.

 

 

 

Figure 1: Block diagram of the pinna landmark estimation process

 

Provided input:

 

A dataset consisting of 3D meshes in the ply format, along with 85 landmarks of the left and right pinna, is provided.

 

The 3D scan provided was captured using the hand-held scanner EinScan pro 2X 2020 around the head and upper torso [2]. Following the scanning process, the scans were then subjected to post-processing to remove any unwanted meshes and fill any holes. Subsequently, the 3D meshes are aligned along the interaural axis [3] illustrated in Figure 2, thereby establishing a consistent anatomical reference frame as outlined below:

 

• The Y-axis runs along the left ear canal to the right ear canal entrance.
• The X-axis is from the back of the head to the front of the head, passing the tip of the nose.
• The Z-axis runs vertically towards the top of the head.
• The center of the head is defined by the intersection of the above-defined X, Y, and Z axes.

 

Figure 2: Coordinate systems 3D mesh [3]

 

This anatomical alignment ensures that all annotations are made consistently across subjects.

 

The annotated landmarks comprised 4 distinct pinna contours:

 

1. Outer helix (25 points)

2. Concha outline (30 points)

3. Inner helix (20 points)

4. Superior antihelix (10 points)

resulting in a total of 85 landmark points, as illustrated in Figure 3 below.

 

 

Figure 3: Pinna Contours

 

The landmarks are constituted by the above-listed pinna contours arranged in an 85 x 3 matrix.

 

Expected output:

 

Participants are required to provide 85 landmarks representing the shape listed above, with separate landmarks for the left and right pinna.

 

Evaluation:

 

The model will be evaluated on an undisclosed set of subjects. The performance of this task is evaluated by comparing the predicted landmarks with the ground truth landmarks.

 

Given a set of 𝑁 ground truth landmarks for one ear of subject j,

 

 

and a set of 𝑁 predicted landmarks,

 

where  contain the coordinates of the  landmark in each set. 

 

The mean Euclidean distance for a single set of landmarks is computed by:

 

The overall performance is then computed by averaging those distances across all M subjects and all ears of the hidden test set:

 

with 

 

Dataset:

 

We provide a sample dataset consisting of a single 3D mesh, along with annotated left and right 85 pinna landmarks. Additionally, a notebook example code is included to demonstrate the basic functionality to help you get started quickly with the challenge.

To gain full access to the complete dataset—comprising 3D head and torso meshes for 200 subjects, along with corresponding annotated pinna landmarks—all team members are required to sign and submit a valid Non-Disclosure Agreement (NDA). Access will only be granted after the NDA has been successfully verified.

 

If you have any questions regarding the implementation details for NDA, please refer to the Delivery section on the platform or contact the Huawei team via the challenge platform.

 

Reference:

 

[1]. https://en.wikipedia.org/wiki/Outer_ear

[2]. Engel, Isaac & Daugintis, Rapolas & Vicente, Thibault & Hogg, Aidan & Pauwels, Johan & Tournier, Arnaud & Picinali, Lorenzo. (2023). The SONICOM HRTF Dataset. Journal AES.

[3]. Kahana Yuvi & Nelson Philip, “Boundary element simulations of the transfer function of human heads and baffled pinnae using accurate geometric models”, Journal of Sound and Vibration. 2007.