# keypoint list?

I would like to try some structure from motion and related epipolar geometry algorithms. Is there a way to get the list of matched keypoints instead of the kptmatch object?

Right now there isn’t, but, we can add it.

What would you like to see implemented?

I would think it would be something analogous to a list of cv::Keypoint.

My end goal is to do camera motion estimation. if I can publish the keypoint information across one of the serial ports, I can also run other post-processing

If the Cam 7 is up to the task, I think a trajectory estimator would be a high value add.

https://docs.opencv.org/3.1.0/d5/dab/tutorial_sfm_trajectory_estimation.html

Hi,

I’ll add this to the next release, however note each keypoint consumes a lot of memory (the descriptor alone is 32 bytes), converting a list of keypoints to Python objects could make the camera run out of memory, so this might not be very usable until our next camera is released.

You could do this without accessing individual keypoints, using the delta of kptmatch objects centers (cx2-cx1, cy2-cy1).

I also tried this before using optical flow, it worked very well (I draw a map of my movement through the room) I’m attaching a couple of (very rough) scripts I used (one for the camera, the other for host side, it uses pygame to draw the map).
optflow.py (1.35 KB)

These were useful examples, thanks.

In my case I’m travelling down a hallway with a fair number of wall features plus occasional obstacles. I’m not localizing/mapping, so the displacement information isn’t as useful as position relative to the side walls or obstacles. I currently have my MV 7 pointing forward, and I’m using find_displacement with two small rectangular areas on the left and right. I subtract their absolute values to find differential forward minus common spin.
diff = math.fabs(delta_x_l) - math.fabs(delta_x_r)

This works fairly well for centering movement down a hallway until there is a junction or obstacle. I suspect I could get more useful information with furthur field segmentation, but with keypoint matches I can go directly to more established sfm algorithms.

There could be some reduction in the data. I don’t need the full keypoint info, just the matched point locations in the image. Two lists of point locations where the same indexes represent the match should be good.

On a somewhat related issue, most of what I am doing will be memory intensive so if there is an option on the next design iteration for a larger ram capacity at a higher price, I would be interested in that.

The H7 model that will come out next year will have 1 MB of RAM. I think about 256KB will be used for the MP heap.

Hi, this is now implemented in the development firmware. The keypoints match object has a list of matching keypoints indices (a list of tuples) you can use to subscript keypoints arrays. For example:

``````for m in match.match():
print(kpts1[m[0]], kpts2[m[1]])
``````

The keypoint is a tuple of (x, y, score, octave, angle). Note you’ll have to download and flash the firmware (or wait for the release):

https://github.com/openmv/openmv/raw/master/firmware/OPENMV3/firmware.bin

Hi, I have a problem with matching the key point
I have several items I want to identify with openmv
But when the file size is more than 70 kilobytes
ide returns this error

I think it’s because of the low memory performance of the camera
Can we use external memory as executable(ram) memory?
New folder.rar (55.1 KB)

Yes the camera has limited amount of memory. Try to lower the number of keypoints (increase the thresholds).

It’s possible but the camera doesn’t have an external ram on board.