search

frotms / Curve-Text-Rectification-Using-Pairs-Of-Points

A way to rectify curve text images using spatial transformer by pairs of points.
Apache License 2.0
30 stars 7 forks source link

How can i get the theory？ #1

Open xudongwang0828 opened 3 years ago

xudongwang0828 commented 3 years ago

What a fantastic work！But i don't know the theory, can you tell me the blog or paper?Very thanks!

frotms commented 3 years ago

@xudongwang0828 I take it as a fisheye distortion problem. 3D modeling with a virtual camera and fisheye camera model may rectify curve text in image. For details, see the Reference of readme.md. And simple method based on Homography for higher FPS will be updated lately.

xudongwang0828 commented 3 years ago

@xudongwang0828 I take it as a fisheye distortion problem. 3D modeling with a virtual camera and fisheye camera model may rectify curve text in image. For details, see the Reference of readme.md. And simple method based on Homography for higher FPS will be updated lately.

Thank you for your reply.I read the code, and it maybe work for 14 points, but if points more than 14, result looks bad.Do you know the reason?

frotms commented 3 years ago

@xudongwang0828 I take it as a fisheye distortion problem. 3D modeling with a virtual camera and fisheye camera model may rectify curve text in image. For details, see the Reference of readme.md. And simple method based on Homography for higher FPS will be updated lately.

Thank you for your reply.I read the code, and it maybe work for 14 points, but if points more than 14, result looks bad.Do you know the reason?

In calibration mode, it failed when points could not satisfy fisheye distortion condition, and the ret of cv2.calibrateCamera is large.

frotms commented 3 years ago

@xudongwang0828 I take it as a fisheye distortion problem. 3D modeling with a virtual camera and fisheye camera model may rectify curve text in image. For details, see the Reference of readme.md. And simple method based on Homography for higher FPS will be updated lately.

Thank you for your reply.I read the code, and it maybe work for 14 points, but if points more than 14, result looks bad.Do you know the reason?

In calibration mode, it failed when points could not satisfy fisheye distortion condition, and the ret of cv2.calibrateCamera is large.

@xudongwang0828 There is a bug for more than 14 points.

world_coord = np.ones((14, 3)) * 0