Open pwshen opened 2 years ago
@pwshen Good point! It is correct.
In short, the transform of resize and crop is learned inherently by the network.
In fact, what we always want is the lane coordinates in the original image space. In this way, we have two options:
The two options' learning difficulties are the same. For example, if a network can learn y=f(x), then it can definitely learn y=3*f(x).
https://github.com/cfzd/Ultra-Fast-Lane-Detection-v2/blob/master/data/dali_data.py#L160
At data/dali_data.py from line 152-160, points are transformed with image warp, at this step points are aligned with lanes on image
After warp affine (from line 163-176), the image is resized and cropped, but there's no further transform to the points, shows that the points are not at the correct spots of the image.
Is this a bug, or you did this with some purpose?