Refactor evaluation flow #309
Conversation
| if isinstance(sample, np.ndarray): | ||
| result = np.round(sample * shape[:1:-1]) | ||
| else: | ||
| result = [np.round(s * shape[:1:-1]) for s in sample] |
There was a problem hiding this comment.
shape[:1:-1]) --> shape[::-1]. otherwise, it will be wrong if scale_h, and scale_w are different.
There was a problem hiding this comment.
the shape's values are [h, w, scale_h, scale_w]. shape[:1:-1] takes the last 2 values (which are scale_h and scale_w) and reverse them to match x, y order of data.
| self._scale_fields = rescale_fields | ||
|
|
||
| @staticmethod | ||
| def _scale_sample(sample: Union[List[np.ndarray], np.ndarray], shape: np.ndarray): |
There was a problem hiding this comment.
the name sample is easy to get confuse with data sample. but the input here are polygons, curves, or boxes for text region representation. replacing sample with field or result will be intuitive, e.g., _rescale_result(...)
| def _scale_sample(sample: Union[List[np.ndarray], np.ndarray], shape: np.ndarray): | ||
| # shape: [h, w, scale_h, scale_w] | ||
| if isinstance(sample, np.ndarray): | ||
| result = np.round(sample * shape[:1:-1]) |
There was a problem hiding this comment.
Either we np.clip the rounded coordinates in origin_width, orgin_height to avoid being out of the image. Or we don't clip and still keep float type here, then np.clip all at one in the end.
|
related to this pr: #316 |
| import cv2 | ||
| import numpy as np | ||
| from shapely.geometry import Polygon | ||
| import mindspore as ms |
There was a problem hiding this comment.
Line#236 and Line#242 need this import.
Thank you for your contribution to the MindOCR repo.
Before submitting this PR, please make sure:
Motivation
det_postprocess.pyto match the project's default data handling flow.