add OCR Decoding support - WIP #113
Conversation
| class_mapping: Dict[str, int], | ||
| **_, | ||
| ) -> np.ndarray: | ||
| text_labels = None |
There was a problem hiding this comment.
text_labels can be already set to np.zeros((len(annotations), ann.max_len)) so there's no chance to return None
| @@ -174,6 +174,7 @@ def _load_image_with_annotations(self, idx: int) -> Tuple[np.ndarray, Labels]: | |||
|
|
|||
| uuid = self.instances[idx] | |||
| df = self.df.loc[uuid] | |||
| print(df) | |||
|
|
||
|
|
||
| def validate_text_value( | ||
| value: str, |
There was a problem hiding this comment.
The annotation of value seems to be incorrect.
| @type is_train: bool | ||
| """ | ||
| super(OCRAugmentation, self).__init__() | ||
| self.transforms = A.Compose( |
There was a problem hiding this comment.
Is this a set of some standard augmentations that are usually performed for OCR task or how is this defined?
| ], | ||
| p=0.2 | ||
| ), | ||
| A.Compose( # resize to image_size with aspect ratio, pad if needed |
There was a problem hiding this comment.
Resize and Normalize are already part of the default augmentations. Resize is always done (you can control if it keeps aspect ratio or not) and Normalize is also appended to list of augmentations (if used by luxonis-train, can be deactivated through config though). So is this needed here?
| @param is_train: True if image is train. False if image is val/test. | ||
| @type is_train: bool | ||
| """ | ||
| super(OCRAugmentation, self).__init__() |
There was a problem hiding this comment.
Le'ts keep it just super().__init__(). The arguments in super are a relic from python 2.
There was a problem hiding this comment.
Next change will be adding kpt label to the TEXT LabelType, even though the goal is only OCR recognition, on the data side it makes since to create the annotation/LabelType from the begining to support kpt annotations
What is meant by this? We have the LabelType.KEYPOINTS already. We also plan to support nested annotations, so I think the final form for OCR + keypoints would be TEXT and KEYPOINTS nested within a BOUNDINGBOX
| def set_global_metadata(self, metadata: Dict[str, Any]) -> None: | ||
| self.global_metadata = metadata |
There was a problem hiding this comment.
Due to GCS datasets, I think we need a way to persist this via storage instead of just memory? Perhaps we could use the existing datasets.json or metadata folder?
This is a WIP to add CTC OCR recognition/decoding
Conformity to contribution guidelines will be fixed before closing
Next change will be adding kpt label to the TEXT LabelType, even though the goal is only OCR recognition, on the data side it makes since to create the annotation/LabelType from the begining to support kpt annotations