Corner Detection

This site is best viewed in a modern browser with JavaScript enabled.

rbuss

We have a leica 3d. A big issue is finding corners as the start or end point of a measurement. Sometimes inside corners are not ‚sharp‘ being filled with paint, plaster or spachtel. Outside edges are also hard to hit. It would be good to have the system localize or find the corner. Since you have a 3D view you want to pick the start point based on the intersection of the planes. I imagine having a circle that you turn the phone so that the corner is in the circle. When it is identified if can change color and place the point at the intersection. An okay confirms the the point is correct and set up the search for the next point.

yang

Hi rbuss,

Thanks for your information! If I understand it correctly, you want an automatic corner point detection. There are two ways to do it:

use Apple or Google's AR frame to do the plane detection, it's all based on software and camera.
Pros: completely automatic;
Cons: not accurate.
we can add a feature to locate the corner point using Arrim ONE's laser beam. Based on theory of three points define a plane, you click three times on a wall, and system will generate a virtual plane. Once you specify three planes, the corner point is accurately located.
Pros: super accurate
Cons: more clicks needed

Let me know what you think of these two ways. We can further discuss about it.

rbuss

yang Hmmm. There are different things at issue here. First lets accept that buildings don‘t accurately follow the rules of geometry - walls are not pure planes but can be twisted and bowed. At what point do we accept three (or four) points to define a plane and when do we need extra points to define the wall? I suggest that five points are good for 3 m walls, and more are necessary for longer walls.

Measuring a window niche has four side planes and the window itself inside a large wall. The side planes aren‘t too sigificant since then tend towards 90 degrees and are short. The built intention is usually clear. The window itself should be parallel to the main wall. If it is not, then it is „interesting“ because you might want to correct it when replacing the windows and need to deal with the window sill that no longer fits, because it was cut crooked to fit the window.

I think part of the isue needs to be answered with what you are trying to measure. If my job is the replace the windows then I want to know how big the windows are and probably don’t want to know that the window is crooked. As the architect I‘m more interested in knowing that the window is crooked than the exact window dimensions - the window supplier needs the millimeter exact dimensions. I just need accurate enough dimensions to get a price.

If I am honest about what I want, I would like a model of all the planes measured in 3d. You take the camera and detect the planes and then ask the user to laser the key points with the option to ignore things - no need to measure this window if I know I am going to close the opening. For long walls you would target extra points to get an accurate model even if the wall is bowed and twisted. You show the wall with target points on the screen and the user sets the laser on the point and verifies it. As the user moves around this gets complicated but it is easier to ensure that surfaces are measured accurately on all sides. I accept that this is not version 1.0

For V1.0 with just point to point measurements, how about a corner detect function that uses the camera to identify a corner and the laser to verify the corner position. In the case of a window - desired length and width - the camera (AR whatever) sees a corner and generates a grey transparent circle over the corner. This should be an area about 2 cm in diameter. The user lasers the corner and then you correct the start poition to the actual intersection point. The laser measured point is close to the actual corner - from 0 to 1 cm. You improve this with plane detection to get the best start point. The end point is done the same way.

Edges can be done similarly. Detect two planes and measure from the edge. This brings up another „feature“. When measuring edge to edge, I may need to know that both points are at the same height. Could you place a dashed level line in the display so points can be taken at the same height? I can see other uses for this function. First are things level, and second what is the angle of a ceiling relative to a level line. Vertcal lines are equally useful.

Without a tripod I don’t know how accurate you can measure anyway. Both the motion to swing from one corner of a room to the other, and position of the person leave room for error. Comparison measurements with and without tripod will determine if a tripod is always necessary.

rbuss

yang Rereading this, I still haven’t said A or B. I think 9 points to identify a corner is too many and 3 points for a plane that covers a large distance too few. If we can assume that all planes meet in a corner at 90 degrees (and that the user needs to make this call) then is it possible to identify the corner with three points? These points need to be within say 50 millimeters from the corner to be useful. Again it is a combination of camera AR und laser measurements.

Other than 9 points being tedious for a corner, there is an issue of accuracy. You have plus minus 1,5 mm. If this is for every measurement, then the plane detection with 3 measurements all within 10 cm from the corner is not going to be too accurate. Whatever method is chosen, it needs to reflect the accuracy of the system/process. For a window niche that is only 10 cm deep, I can‘t generate a plane with larger distances. Without a tripod, there will be movement between measurements. So what is the math to an optimal solution? Hitting the corner with Xmm accuracy versus AR corner versus AR corner with one point versus AR corner with three points versus 9 point laser corner. The improvement in accuracy needs to justify more work (programming and use). I can measure a window (length x height) with three points if I have the corners an four if I take the middle points. At the midpoints I don‘t need to resolve rounded corners. The corner detection should help by making it easier to measure multiple things from the same point accurately. If it is too complicated, then people will take multiple measurements with a loss of accuracy, or duplicate measurements. (Wall segment plus window plus wall segment = complete wall length. - How much variance is there between the two.)

Where can I send photos of corners?

yang

You can post the photos in your reply or send it to hengxing.yang@arrim.net

The dashed leveling line is very useful feature. We will add it in immediately. Thanks for your suggestion!

Back to the conner point topic. You made very solid point about the issue of defining a wall by 3 points. Use camera to identify a corner and the laser to verify the position could be a good solution. However. AR/camera detection has a big technical constraint - only objects with pattern/texture can be recognized. It won't work for white walls or single color background. That's why all AR applications are demonstrated on a wooden table. We uses laser instead of AR detection to get the accuracy and reliability. Corner detection based on AR detection will have pretty bad experience. We are working on this issue, may try it with other technology. We will keep you updated.