I am trying to get the plant as it is by segmenting everything else in the region, but I am unable to pinpoint the boundaries of the colors forming the plant:
It’s a straightforward task that I am using the following code to achieve:
import numpy as np
# specify the path to the video
vidcap = cv2.VideoCapture('/home/bla/Desktop/3d_reconst/data/2018/02/22/1/MVI_0590.MOV')
# read the first frame
success,image = vidcap.read()
# so that we know the number of frames
count = 0
# do it for the entire stream, as long as the video plays, success flag is for this purpose
# switch to HSV color space
hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
#specify the lower and upper boundaries for the color segmentation
low_brown = np.array([10, 100, 20])
high_brown = np.array([20, 255, 200])
# take the foreground with the specified boundaries
mask = cv2.inRange(hsv, low_brown, high_brown)
# apply it on the original frame and get its original form
res = cv2.bitwise_and(image,image, mask= mask)
# save frame as JPEG file
cv2.imwrite("frame%d.jpg" % count, res)
success,image = vidcap.read()
print('Frame #:', count)
count += 1
What I fail to find though is the lower and upper boundaries for the plant (assumingly it is brownish tones) I am trying to segment. Is there a convenient way to do that? At the moment the intervals are not doing it right, and I get an output where the plant is not segmented at all.
I am also not sure if I should use blurring before I go for segmentation, my input frames are coming from a .MOV file (uses MPEG4 compression) so there is already blur involved in my images.
Any recommendation is highly appreciated.
it’s probably far easier to find the blue background, and then invert the mask
Yeah but then the QR code and the cable on the right side would remain as they are not blueish. I thought it would be easier to use one interval for the plant and eliminate everything else, instead of finding 3 intervals (background, qr code, cable).
select blue as background and invert.
just “blindly” mask away the QR codes on the side. they’re always there, never near the object. define a rectangle covering them, erase that area in the mask.
don’t get your hopes up. picture is blurry, so you’ll get blue bleeding into the object, or the object getting thinner than it is.
the bluescreen isn’t terribly uniform. the more variation it has, the wider the range you have to select, which may take away from color ranges found in the foreground.
if you’re careful with everything, this blur could be dealt with and turned into grayscale transparency information (rather than a hard binary mask). still, the whole process benefits from high resolution and the object being in focus.
the object, the roots, are illuminated by the bluescreen, which is bad. bluescreen must be far enough away that they throw no light onto the foreground. or very little, which can be “drowned out” by illumination on the foreground/object.
I was able to achieve what I wanted. I applied the HSV range thanks to the HSV color picker code I found here. It basically takes an image and lets you play around with min/max HSV values, and you can see the effects in real-time, which is quite nice. After that, it was masking and inverting. I can also crop this just to process the plant root itself. Possibilities are there.
My result is not perfect. However, this is without any preprocessing. I could play around with some properties of the images to perfect this mask.
Try kmeans clustering
press e key to execute
Sorry it’s in French