Polygon output with new API? #18330

You are right — there is no polygon switch on PaddleOCR()
the wrapper forwards only these detection params paddleocr/_pipelines/ocr.py:145-152

box_type is not an init arg at all
it comes from the model config

paddlex/inference/models/text_detection/predictor.py:118

post_op = self.build_postprocess(**self.config["PostProcess"])

:213

box_type=kwargs.get("box_type", "quad"),

and it picks the DB decoder paddlex/inference/models/text_detection/processors.py:531-540

if self.box_type == "poly":
    boxes, scores = self.polygons_from_bitmap(...)
elif self.box_type == "quad":
    boxes, scores = self.boxes_from_bitmap(...)

so flip it on the detector

from paddleocr import TextDetection

det = TextDetection()
det.paddlex_predictor.post_op.box_type = "poly"

res = det.predict("img.png")
print(res[0]["dt_polys"])   # N point polygons instead of 4 point quads

or set box_type: poly under PostProcess in the model dir inference.yml if you want it to stick

do not do this inside the full PaddleOCR pipeline
the general branch is quad all the way paddlex/inference/pipelines/ocr/pipeline.py:111-112

self._sort_boxes = SortQuadBoxes()
self._crop_by_polys = CropByPolys(det_box_type="quad")

recognition would then crop N point boxes with a 4 point cropper

for poly end to end there is already a poly branch — text_type seal pipeline.py:121-122 with SortPolyBoxes and CropByPolys(det_box_type="poly")
SealTextDetection exposes that detector on its own