Tweet
One of the interesting questions related to use of zephyr for measurements of physical objects. This can be broken into 3 sub questions: 1. what type of cameras, 2. how many images and 3.what kind of accuracy that can be expected. I did not how to find the answers. So I setup capture and processing using Zephyr Lite and my own smart phone (Redmi Note 5 Pro with 20 megapixel camera - costs less than $200 - fixed with a selfie stick - $10 - to go around the object of interest. The selfie stick has a wire connecting to the stereo audio jack of the smart phone so that pictures can be taken by pressing the button on the selfie stick. I thought of experimenting with gimbal stabilized selfie stick, but this proved to be too cumbersome as gimbal stabilization added more problems in orienting the camera. Hence I resorted to a simple wired selfie stick that I can hold it steadily while clicking the camera.
The object that I wanted to measure was an Altera (Now Intel) Max10 FPGA development board that measures 76 mm x 89 mm x 8 mm. So I placed the board on a book which I had good texture ("Deep Learning" by Ian Goodfellow, et al, a well known book). I did not worry too much about lighting and just used day light present. I took about 150 pictures all around the board.
Zephyr Lite took more than 4 hours to reconstruct the object with 407,736 vertices (perhaps because of the texture on the cover of the book) and 811,341 triangles. Then I cut with bounding box, aligned with XY plane using up vector using plane button. Then the vertices shrank to 145,494 vertices and 274,343 triangles. The reconstruction pictures are shown here. I used measurement tool to measure the distances and calibrate with known board dimensions.
76 mm had a measurement value of 18.28735 and 89 mm turned out to be 21.48735. If I calculate their ratios, I get a scale factor of either 4.1558782 or 4.14941 which are pretty close. With one of the scale factors, I measure the height of the connector in the board as 7.694 mm (actual value is 8 mm). Hence I had an error of 0.3 mm in any dimension. Taking the smallest dimension of 8 mm, the error ratio is 0.3/8 =3.75% of full scale value. But for longer lengths, say for 89 mm, this error is as close as possible with a ruler. Of course, with a vernier caliper, the accuracy will be much more.
As you can see in the closeup, the reconstruction has holes and some of parts are rounded up. Yet the resolving power is as good as you can measure with a ruler. All done with run of the mill cell phone camera in day light. The image capture took about three minutes.
The object that I wanted to measure was an Altera (Now Intel) Max10 FPGA development board that measures 76 mm x 89 mm x 8 mm. So I placed the board on a book which I had good texture ("Deep Learning" by Ian Goodfellow, et al, a well known book). I did not worry too much about lighting and just used day light present. I took about 150 pictures all around the board.
Zephyr Lite took more than 4 hours to reconstruct the object with 407,736 vertices (perhaps because of the texture on the cover of the book) and 811,341 triangles. Then I cut with bounding box, aligned with XY plane using up vector using plane button. Then the vertices shrank to 145,494 vertices and 274,343 triangles. The reconstruction pictures are shown here. I used measurement tool to measure the distances and calibrate with known board dimensions.
76 mm had a measurement value of 18.28735 and 89 mm turned out to be 21.48735. If I calculate their ratios, I get a scale factor of either 4.1558782 or 4.14941 which are pretty close. With one of the scale factors, I measure the height of the connector in the board as 7.694 mm (actual value is 8 mm). Hence I had an error of 0.3 mm in any dimension. Taking the smallest dimension of 8 mm, the error ratio is 0.3/8 =3.75% of full scale value. But for longer lengths, say for 89 mm, this error is as close as possible with a ruler. Of course, with a vernier caliper, the accuracy will be much more.
As you can see in the closeup, the reconstruction has holes and some of parts are rounded up. Yet the resolving power is as good as you can measure with a ruler. All done with run of the mill cell phone camera in day light. The image capture took about three minutes.
Comment