Processing Time of pan_pp.det_head.get_results() [ms] (Used data: ICDAR2015 [dataset])
Original Source: 118.113
Improved Pixel Aggregation Source: 112.645
It was confirmed that pixel aggregation (pa.pyx) is dominant during the post-processing process for images with variously distributed strings.
Therefore, I modified the cython code to improve speed and parallel processing.
The result is a 4.629% improvement based on the ic15 dataset, but a huge speedup can be expected if there are numerous labels. (Based on the dataset I used (documents), it's about 100 ms shorter.)
Also modified the input of cv2.resize to solve the error below.
label = cv2.resize(label, (img_size[1], img_size[0]),
score = cv2.resize(score, (img_size[1], img_size[0]),
cv2.error: OpenCV(4.7.0) :-1: error: (-5:Bad argument) in function 'resize'
> Overload resolution failed:
> - Can't parse 'dsize'. Sequence item with index 0 has a wrong type
> - Can't parse 'dsize'. Sequence item with index 0 has a wrong type
Dear Mr. Wang
Processing Time of
pan_pp.det_head.get_results()
[ms] (Used data: ICDAR2015 [dataset])It was confirmed that pixel aggregation (pa.pyx) is dominant during the post-processing process for images with variously distributed strings. Therefore, I modified the cython code to improve speed and parallel processing. The result is a 4.629% improvement based on the ic15 dataset, but a huge speedup can be expected if there are numerous
labels
. (Based on the dataset I used (documents), it's about 100 ms shorter.)Also modified the input of
cv2.resize
to solve the error below.Sincerely, Hyogeun Oh