Repo for DroneDeploy's camera localization challenge.
Run camera_localization.py to get camera pose estimates for each image in the root folder.
First, load in and downsample the image by a factor of 10, to save time. Also, load the pattern and threshold the image.
Locate the QR code within the image.
- Use cv2.findContours to find the boundaries of the piece of paper within the image.
- Identify the black pixels with the minimal/maximal x/y coordinates within these borders to get the corners of the QR pattern.
Estimate the camera pose, using the corners of the QR pattern.
- My implementation of SGD and numpy's fsolve both failed to converge to a good answer.
- Use a guided random walk with decreasing step sizes through time.
- At each time step, propose a random step.
- Accept the step if it brings us closer to the goal.
- Definitely an area for improvement.
Simulate the view from four rotations about the z-axis.
- Rotating the camera 90 degrees about the z-axis will still produce the same corners.
- Run camera_simulation.py for an example of this.
- Find the best rotation by comparing the simulated image to the actual image.
Plot the camera coordinates and their image planes!
The x-coordinate increases to the right, the y coordinate increases moving down, and the z coordinate increases out of the screen.
| Image | x | y | z | pitch | yaw | roll |
|---|---|---|---|---|---|---|
| IMG_6719.JPG | -18.7 | 10.0 | 68.0 | 72 | 237 | -7 |
| IMG_6726.JPG | 10.9 | -10.3 | 63.5 | 105 | 239 | 19 |
| IMG_6727.JPG | 26.2 | 3.4 | 84.4 | 111 | 274 | -39 |
| IMG_6722.JPG | -25.3 | -5.6 | 78.8 | 69 | 274 | -14 |
| IMG_6724.JPG | 0.5 | 0.8 | 92.0 | 85 | 237 | 28 |
| IMG_6721.JPG | 29.4 | -11.8 | 67.8 | 114 | 259 | -41 |
| IMG_6723.JPG | -27.1 | -3.1 | 82.5 | 70 | 262 | -44 |
| IMG_6725.JPG | -0.1 | -0.2 | 43.7 | 93 | 216 | -25 |
| IMG_6720.JPG | -32.1 | -2.5 | 79.1 | 70 | 264 | -77 |