Improved SCRFD decoding. #21
Conversation
| @@ -98,7 +98,7 @@ def create_detection_message( | |||
| raise ValueError(f"keypoints should be list, got {type(keypoints)}.") | |||
| for pointcloud in keypoints: | |||
| for point in pointcloud: | |||
| if not isinstance(point, Tuple): | |||
| if not isinstance(point, Tuple) and not isinstance(point, List): | |||
There was a problem hiding this comment.
Nitpick: Is there a better name for a variable than pointcloud. As I understand keypoints are a list of (n,2) keypoints for specific object? Maybe:
for object_keypoints in keypoints:
for point in object_keypoints:
....
Feel free to suggest other names though.
There was a problem hiding this comment.
Hmm yeah good point. I agree with the proposed naming. Added in a6b3b02.
| score_threshold=self.score_threshold, | ||
| nms_threshold=self.nms_threshold, | ||
| ) | ||
| detection_msg = create_detection_message( |
There was a problem hiding this comment.
In line to PR #19 we should also add timestamp to the message.
Should we make this part of the create_detection_message() function already? (Can be a separate PR)
There was a problem hiding this comment.
Yea, added it when you were reviewing. Hmm I was thinking that it would be best to add it to the creator functions but we need output to get the timestamp. Since it is only one line of code we leave it here? Because otherwise we need to propagate output to every creator function.
This PR implements better SCRFD decoding with keypoints. It follows the official decoding procedure. Decoding is moved to the utils files. Some small validation changes are made in detections creators and message class @jkbmrz so it accepts also keypoints as list and not only tuple. This is for easier conversion from numpy array (going only with
np_array.tolist(). This shouldn't break YuNet parser.Some additional parameters are added to the parser e.g. input_size, feature_strides etc. because if we add different scrfd model (lighter, heavier) in the future we only adjust those params.
Note that NMS function is added here. We should address using one NMS for all parsers where posssible in the separate task.
There is also support for RVC4 decoding so it works for both platforms.