Project Aria Glasses 2D Image Coordinate System Conventions

For any provided camera intrinsic calibration value, we use the convention that the color value of a pixel with integer coordinates $(u,v)$ is the average color of the square spanning from $(u-0.5,v-0.5)$ to $(u+0.5,v+0.5)$ in continuous coordinates.

This is visualized in the Figure 1, and has the following important consequences:

• Checking in bound: A pixel $(u,v)$ is considered to be in bound if $-0.5\leq u<W-0.5$ and $-0.5\leq v<H-0.5$.
• Interpolation: In bilinear interpolation, a point (u,v) can be interpolated of all four neighboring integer-valued pixel coordinates are in-bound. That requires $0 \leq u \leq W-1$ and $0 \leq v \leq H-1$.
• Image down-sampling: When downsampling images by a factor of $s$, every $s \times s$ pixel are squeezed into a single pixel. For example, the intensity at pixel $s \times s$ in the scaled image accounts for all the photons collected in the area $[-0.5,s-0.5]\times[-0.5,s-0.5]$ (i.e. column $0$ to $s - 1$, and row $0$ to $s-1$ in the discrete coordinate) in the original image. In order to keep this assumption valid, the re-scaled point $p_\text{scaled}$ not only needs to scale from the corresponding point in the original image $p_\text{original}$ but also accounts for the $(0.5,0.5)$ translation accordingly by $p_\text{scaled} =s (p_\text{original}+0.5)-0.5$

Figure 1: 2D Image Coordinate System Conventions

Go to the Project Aria FAQ for more calibration information and resources.