SurfaceSight: FREQUENTLY ASKED QUESTIONS
How unique are object contours?
This is perhaps the biggest “ah ha” of the work and a central contribution. It turns out that 2D object contours are surprisingly distinctive, especially if one selects a context of use (e.g., kitchen objects).
Training on a single object view is insufficient (will overfit; poor recognition). That is precisely why we trained on all object orientations (from the paper: “it is important to expose all sides of an object to SurfaceSight during training […] in the training phase, we captured data for each object … different positions, different angles”). To evaluate this, we also placed objects in random orientations and positions, and this is the accuracy we report (~94%).
Is there a relationship between BMI and accuracy on user angle estimation?
This is why we ran a 14 person user study as opposed to a performance evaluation. It offers some initial clues as to the performance across users. We collected self-reported height and weight in our demographics intake form, and from this we can report that we had 8 normal weight, 4 overweight and 2 obese users (NIH BMI scale). There is no correlation in weight and accuracy.
Why not use a camera?
360 degree cameras are powerful, but must deal with variable lighting, require more compute to process, and have privacy implications in e.g., the home. Contours from LIDAR have known real-world scale, unlike cameras that must compute scale-invariant features.
Depth cameras could conceivably be made 360, or several could be used in concern (large/expensive). Although there are many technologies in use, very few are millimeter accurate, which is what we need for accurate contours.
Radar tends to be less directional, with most wavelengths either passing through or scattering on objects we included (plastic bowls or metal rice cookers). Sonar too is not directional/precise enough.
Yes, LIDAR has “primitive curves” with “very little features”, but they are also extremely precise, capturing even small differences in contour shape/angle, which enables our approach.