torchocr weights

Pretrained checkpoints for torchocr, a PyTorch-native, torchvision-style OCR library. You normally never download these files by hand: each one backs a member of a *_Weights enum, and torch.hub fetches it on first use and checks the SHA-256 prefix embedded in the file name.

from torchocr import DBPostProcessor
from torchocr.models import DBNet_ResNet18_VD_Weights, dbnet_resnet18_vd

weights = DBNet_ResNet18_VD_Weights.ICDAR2015
model = dbnet_resnet18_vd(weights=weights).eval()
preprocess = weights.transforms()
postprocess = DBPostProcessor(**weights.meta["postprocess"])

Checkpoints

Page orientation (0, 90, 180 or 270 degrees), scored on the 4999 DocLayNet test pages, each fed upright and turned three ways (references/orientation/evaluate.py):

⁴ The threshold was chosen on DocLayNet validation and applied once to test. Weakest on manuals (0.973) and laws (0.979). Meant for document pages, not scene photos.

¹ Line-level detectors converted from PaddleOCR. ICDAR-2015 scores word boxes, so adjacent words merged into one line count as misses: these numbers compare checkpoints, they are not the accuracy you will see on documents. ² 2077 alphanumeric care words of ICDAR-2015 test cropped from the full images with torchocr.ops.crop_quads on their ground-truth quads, case-insensitive (not the official Task 4.3 crops). Axis-aligned crops of the same words score 0.579. ³ Decoded with its declared CTC beam search (width 10) and 6-gram character language model; 0.750 with greedy decoding (0.757 for the MJSynth checkpoint). Same ground-truth-quad protocol; Total-Text curved words use their minimum-area rectangle.

End to end (OCRPipeline.from_pretrained, image → text; a word counts when it is detected and read correctly, case-insensitive, generic vocabulary; references/end2end/evaluate.py), with DBNet_ResNet18_VD_Weights.ICDAR2015 and each scene recognizer, min_text_score=0.8:

Without min_text_score, the v0.3 recognizer scores 0.621 / 0.566 / 0.711.

min_text_score=0.8 drops regions read with low confidence; the value maximizes end-to-end hmean on the ICDAR-2015 training split. The converted PaddleOCR weights reach 0.219 on ICDAR-2015 and 0.17–0.18 detection hmean on FUNSD, whose word-level boxes do not match their line-level output.

from torchocr import OCRPipeline, load_image

pipeline = OCRPipeline.from_pretrained(
    "DBNet_ResNet18_VD_Weights.ICDAR2015", "CRNN_ResNet34_VD_Weights.MJSYNTH_ICDAR2015_TOTALTEXT_MLT19", device="cuda"
)  # the recognizer's beam search + language model is set up automatically
pipeline.min_text_score = 0.8
doc = pipeline(load_image("photo.jpg"))  # doc.text, doc.polygons

Evaluation protocol

Detection: official ICDAR-2015 IoU protocol (one-to-one greedy matching at IoU > 0.5, ### don't-care filtering) on rotated quads, implemented by torchocr.metrics.DetectionHmean and verified identical to PaddleOCR's DetectionIoUEvaluator. Pre/post-processing is exactly weights.transforms() and weights.meta["postprocess"]. Every number is reproducible with:

python references/detection/evaluate.py --root <icdar2015> --weights <enum>
python references/recognition/evaluate.py --root <icdar2015> --weights CRNN_ResNet34_VD_Weights.PPOCR_SERVER_V2

The ICDAR2015 checkpoints

All were fine-tuned by torchocr on training splits only (ICDAR-2015, plus Total-Text and MLT-2019 for the recognizer noted below); reported weights are the last epoch and no checkpoint or hyper-parameter was selected on the test set.

• DBNet_ResNet18_VD_Weights.ICDAR2015: PPOCR_SERVER_V2 fine-tuned 300 epochs with the recipe below; box_thresh = 0.5 maximizes train-split hmean (0.794 with PaddleOCR's 0.6).
• CRNN_ResNet34_VD_Weights.ICDAR2015: PPOCR_SERVER_V2 fine-tuned 30 epochs on the 4467 legible training words, cropped with crop_quads from jittered ground-truth quads (references/recognition/train.py). It keeps PaddleOCR's Chinese + Latin charset but saw only Latin scene text, so Chinese accuracy is no longer guaranteed.
• CRNN_ResNet34_VD_Weights.ICDAR2015_TOTALTEXT_MLT19: PPOCR_SERVER_V2 fine-tuned 10 epochs (bf16) on 66k Latin scene words (ICDAR-2015 train ×2, Total-Text train, MLT-2019 Latin crops). Its language model (charlm6_scene-883dd7cd.json) is a case-insensitive 6-gram character model built from those training transcriptions only; beam width and LM weight were chosen on Total-Text training words.
• CRNN_ResNet34_VD_Weights.MJSYNTH_ICDAR2015_TOTALTEXT_MLT19: the checkpoint above fine-tuned 6 more epochs (bf16), each drawing 370k words at random: the 66k real words once on average plus 300k of the 7.2M MJSynth synthetic words (14 undecodable images skipped). Same decoder and language model.
• DBNet_MobileNetV3_Large_05_Weights.ICDAR2015, described next.

DBNet_MobileNetV3_Large_05_Weights.ICDAR2015

PPOCR_V3_EN fine-tuned for 300 epochs on the 1000 ICDAR-2015 training images (references/detection/train.py, PaddleOCR's augmentation recipe, 640 px crops, AdamW 1e-3 with warmup + cosine, ~80 min on one laptop GPU). The reported weights are the last epoch: no checkpoint was selected on the test set. box_thresh = 0.45 maximizes hmean on the training split (references/detection/calibrate.py) and was applied once to the test split; with PaddleOCR's default box_thresh = 0.6 the same checkpoint scores 0.688.

Intended inputs: incidental scene text in street-level photos, Latin script, evaluated at 736 × 1280. Expect lower recall on dense documents and on non-Latin scripts.

Provenance and licenses

• PPOCR_* checkpoints are converted from PaddleOCR releases (Apache-2.0) using the build-time converters in torchocr's scripts/, whose recipes are adapted from PaddleOCR2Pytorch (Apache-2.0). The PP-OCRv3 files hold the Student network that PaddleOCR exports for inference. PPLCNet_x1_0_Weights.PPOCR_DOC_ORIENTATION is PaddleOCR's PP-LCNet_x1_0_doc_ori (Apache-2.0), converted by scripts/convert_paddle_pplcnet.py. It is evaluated on DocLayNet (CDLA-Permissive-1.0), not trained on it.
• The ICDAR2015 checkpoint derives from PPOCR_V3_EN (Apache-2.0) and was fine-tuned on the ICDAR 2015 Incidental Scene Text training set, which its organizers release for research use. Check that your use complies with those terms.
• CRNN_ResNet34_VD_Weights.MJSYNTH_ICDAR2015_TOTALTEXT_MLT19 was additionally trained on MJSynth (Jaderberg et al., 2014), released by the Oxford Visual Geometry Group for research use.