Image reprojection

My friend Brien is a bit of an artist and a bit of a tech nerd so he thought it would be a great idea to make a timelapse video of him drawing. So he put the CHDK firmware on his Canon S90 camera and set it up to take an exposure every 2-3 seconds. He reprojected the drawing portion of the image back into a rectangle and made a video out of it. He managed to do the original video using only ImageMagick command line tools. You can watch his Timelapse Drawing 1 on YouTube.

That worked out pretty well, but we thought it would also be cool to make a video with both the original image containing Brien and the extracted drawing image shown side by side. This would allow one to see the drawing progress and the artist at work.

I thought this would be a good chance for me to do a little image manipulation in Python, so I fired up my IPython notebook and make use of OpenCV.

I started the process off by working through and validating each step I wanted to perform, beginning with standard imports and reading in the sample image.

Now that I have read in the sample image, I want to see what it looks like. Note that below I place a ; at the end of my call to imshow(), this supresses the return value that the IPython Notebook automatically prints out for every cell. We only care to see the image itself. I'll do this every time I use imshow().

Defining Box

We are interested in extracting the region of this image that contains the drawing. To do that I need to identify the corner coordinates of the drawing. I was able to roughly pick them out from the image above but actually I wanted to get slightly more accurate points. So I opened the image separately on the IPython command line and showed the image with plt.imshow() so I could zoom in and get more accurate coordinates for the corners. Recent versions of IPython have much improved interactive widgets but I am unaware of anything that works as easily as Matplotlib's imshow().

I use the coordinates found above to define a NumPy float32 array that contains the original coordinates of the image corners: orig_pts. Then I define the coordinates I want the original points to be mapped to: dest_pts. The ordering here doesn't matter except that the orders must match in both arrays. The transformation we define later will use both the orignal and destination coordinates as the basis of the transformation, so the first element of orig_pts will be transformed to the first element of dest_pts, the second element of orig_pts will be transformed to the second element of dest_pts and so on.

We will want the order of these coordinates for visualizing the region extracted later, so the order is: top left corner, top right corner, bottom left corner and bottom right corner.

Verifying Selected Box

Now, lets verify that the region we've defined by corner coordinates is actually the region we are interested in. Below, I make a copy of the drawing, then use cv2.line() to draw the box defined by the coordinates.

Transforming and Extracting Drawing

The box above looks good, so I proceed with defining the transform using cv2.getPerspectiveTransform() and then use that transform with cv2.warpPerspective to project and extract an image of just the drawing.

That turned out to look pretty good. The left hand side is well defined but I was off a bit on the top right corner and the right hand edge appears to not be straight anyway. Also, since this is just one of 3000 images, I am not going to fine tune anything based on this single image selection. I think I will call this good enough.

Resizing Original

Since we want to put these two images back together and side by side, they have to have the same height. So, below I resize the original image with cv2.resize(), to be 2000 pixels high and 2667 pixels wide to preserve the aspect ratio.

Combining Images

At this point, we have the extracted drawing and the resized original image brien_drawing_smaller, we just need to put them together and check out the result. Since these are both just NumPy arrays, I will use np.hstack() to place them side by side. Then I view the result below.

Cropping and Scaling the Final Image

The resultant image looks great, but the aspect ratio is suboptimal for generating a video. Below we use the shape of the combined image and the desired aspect ratio, 16:10, to compute then generate the final_image. I am going to crop out the left hand side of this image since there is little action on that edge, so first off, I will compute the left hand starting point of our final_image.

Now, with the starting_column identified, I will display the cropped image.

Wrap Up

So the image above is an example of what we'd like a single frame to look like. The next step is to clean this code up, and wrap it up in a function and then run that on all the input images. I'll do this work later and it might result in a future blog post.

If you're wondering about how I drafted this blog post inline with my Python code then you should check out the IPython notebook. Its great for exactly this type of thing. For a broader sense of what it does you can check out the IPython presentation I gave to DesertPy in the spring of 2014.