HB16888/MonkeyOCRv2_det_dptext

MonkeyOCRv2 Detection (DPText-DETR)

This repository provides the DPText-DETR text detection experiments from the MonkeyOCRv2 paper. The visual encoder from MonkeyOCRv2-AS (ViTAEv2-S, 21M parameters) is integrated into DPText-DETR as a drop-in detectron2 backbone. The last three ViTAEv2 stages (strides 8/16/32) are exposed as res3–res5 and feed the standard deformable-DETR input projections, so the transformer encoder/decoder and the detection head are unchanged.

Training and evaluation follow the official DPText-DETR protocols on Total-Text, CTW1500, ICDAR19-ArT, Rotated Total-Text and Inverse-Text.

Models and Results

For each benchmark, three visual backbones are compared under identical settings: the original ImageNet-pretrained ResNet-50, the text-specific oCLIP ResNet-50, and MonkeyOCRv2. MonkeyOCRv2 consistently improves F-score across all five benchmarks.

All models are trained directly on the target dataset (no SynthText/MLT pre-training), for 200k iterations with a total batch size of 8, using the positional label form and the rotated training images released with DPText-DETR (*_poly_train_rotate_pos).

Total-Text

Method	P	R	F
DPText-DETR (ResNet-50)	89.6	82.8	86.1
DPText-DETR + oCLIP	87.1	84.5	85.7
DPText-DETR + MonkeyOCRv2	90.9	86.7	88.8

CTW1500

Method	P	R	F
DPText-DETR (ResNet-50)	89.7	82.1	85.7
DPText-DETR + oCLIP	86.3	82.7	84.5
DPText-DETR + MonkeyOCRv2	89.6	88.1	88.9

ICDAR19-ArT

Method	P	R	F
DPText-DETR (ResNet-50)	84.3	67.5	75.0
DPText-DETR + oCLIP	75.1	62.0	67.9
DPText-DETR + MonkeyOCRv2	85.8	71.7	78.1

Rotated Total-Text

Method	P	R	F
DPText-DETR (ResNet-50)	89.4	79.8	84.3
DPText-DETR + oCLIP	87.2	80.8	83.9
DPText-DETR + MonkeyOCRv2	89.7	84.4	86.9

Inverse-Text

Method	P	R	F
DPText-DETR (ResNet-50)	92.1	81.3	86.4
DPText-DETR + oCLIP	90.2	82.1	85.9
DPText-DETR + MonkeyOCRv2	91.8	85.4	88.5

Rotated Total-Text and Inverse-Text are test-only benchmarks: they reuse the Total-Text model above and only change DATASETS.TEST.

Checkpoints

Download the checkpoints from HB16888/MonkeyOCRv2_det_dptext (HuggingFace) or WangXinhan/MonkeyOCRv2_det_dptext (ModelScope):

# HuggingFace
hf download HB16888/MonkeyOCRv2_det_dptext --include "*.pth" --local-dir ./model_weight
# ModelScope
modelscope download --model WangXinhan/MonkeyOCRv2_det_dptext --local_dir ./model_weight

Environment

The reproduced environment uses Python 3.11, PyTorch 2.9.0, CUDA 12.8, torchvision 0.24.0, detectron2 0.6, NumPy 2.4.4, Transformers 4.57.1 and safetensors 0.7.0. The oCLIP baseline additionally needs MMOCR 1.0.1 (MMEngine 0.10.7, MMCV 2.0.1, MMDet 3.1.0). All models were trained on 8 GPUs (NVIDIA GeForce RTX 3090) with SOLVER.IMS_PER_BATCH: 8 for 200k iterations.

Installation

This directory is an add-on on top of the official DPText-DETR release. Run:

bash install.sh            # clones DPText-DETR into ./DPText-DETR and patches it

Pretrained Backbones

cd DPText-DETR

# MonkeyOCRv2-AS visual encoder (for the MonkeyOCRv2 rows)
hf download zenosai/MonkeyOCRv2-AS --local-dir ./pretrained/monkeyocrv2_as

# ImageNet ResNet-50 (for the baseline rows) - from the official DPText-DETR /
# AdelaiDet instructions
mkdir -p ckpts
wget -O ckpts/R-50.pkl https://dl.fbaipublicfiles.com/detectron2/ImageNetPretrained/MSRA/R-50.pkl

# oCLIP ResNet-50 (for the oCLIP rows)
wget -O ckpts/resnet50-oclip-7ba0c533.pth \
  https://download.openmmlab.com/mmocr/backbone/resnet50-oclip-7ba0c533.pth

Datasets

Download Total-Text (including rotated images), CTW1500 (including rotated images), ICDAR19-ArT (including rotated images), Inverse-Text, the polygon json files and the evaluation ground-truths from the official DPText-DETR data preparation links, and organize them as:

datasets/
├── totaltext/
│   ├── train_images_rotate/
│   ├── test_images_rotate/
│   ├── train_poly_rotate_pos.json
│   ├── test_poly.json
│   └── test_poly_rotate.json
├── ctw1500/
│   ├── train_images_rotate/
│   ├── test_images/
│   ├── train_poly_rotate_pos.json
│   └── test_poly.json
├── art/
│   ├── train_images_rotate/
│   ├── test_images/
│   ├── train_poly_rotate_pos.json
│   └── test_poly.json
├── inversetext/
│   ├── test_images/
│   └── test_poly.json
└── evaluation/
    ├── gt_totaltext.zip
    ├── gt_totaltext_rotate.zip
    ├── gt_ctw1500.zip
    └── gt_inversetext.zip

Training

All commands are run from the DPText-DETR root directory, on 8 GPUs.

# ---------- Total-Text (also used for Rot.Total-Text and Inverse-Text) ----------
python tools/train_net.py --config-file configs/DPText_DETR/TotalText_Direct_Rotate/R_50_poly.yaml             --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/TotalText_Direct_Rotate/R_50_oclip_poly_lr1e4.yaml --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/TotalText_Direct_Rotate/mkv2vitae_align.yaml       --num-gpus 8

# ---------- CTW1500 ----------
python tools/train_net.py --config-file configs/DPText_DETR/CTW_Rotate/R_50_poly.yaml             --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/CTW_Rotate/R_50_oclip_poly_lr1e4.yaml --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/CTW_Rotate/mkv2vitae_align.yaml       --num-gpus 8

# ---------- ICDAR19-ArT ----------
python tools/train_net.py --config-file configs/DPText_DETR/ArT_Rotate/R_50_poly.yaml             --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/ArT_Rotate/R_50_oclip_poly_lr1e4.yaml --num-gpus 8
python tools/train_net.py --config-file configs/DPText_DETR/ArT_Rotate/mkv2vitae_align.yaml       --num-gpus 8

Evaluation

Each config already carries the MODEL.TRANSFORMER.INFERENCE_TH_TEST value that reproduces the corresponding row of the tables above, so evaluating on the dataset a model was trained on needs no extra flags:

# Total-Text
python tools/train_net.py --num-gpus 8 --eval-only \
  --config-file configs/DPText_DETR/TotalText_Direct_Rotate/mkv2vitae_align.yaml \
  MODEL.WEIGHTS model_weight/dptext_mkv2vitae_totaltext.pth

# CTW1500
python tools/train_net.py --num-gpus 8 --eval-only \
  --config-file configs/DPText_DETR/CTW_Rotate/mkv2vitae_align.yaml \
  MODEL.WEIGHTS model_weight/dptext_mkv2vitae_ctw1500.pth

Evaluation prints precision / recall / hmean on the copypaste: line, matching the tables above.

Rotated Total-Text and Inverse-Text

These reuse the Total-Text checkpoints and override the test set and the threshold:

# Rotated Total-Text
python tools/train_net.py --num-gpus 8 --eval-only \
  --config-file configs/DPText_DETR/TotalText_Direct_Rotate/mkv2vitae_align.yaml \
  MODEL.WEIGHTS model_weight/dptext_mkv2vitae_totaltext.pth \
  MODEL.TRANSFORMER.INFERENCE_TH_TEST 0.395 \
  DATASETS.TEST '("totaltext_poly_test_rotate",)'

# Inverse-Text
python tools/train_net.py --num-gpus 8 --eval-only \
  --config-file configs/DPText_DETR/TotalText_Direct_Rotate/mkv2vitae_align.yaml \
  MODEL.WEIGHTS model_weight/dptext_mkv2vitae_totaltext.pth \
  MODEL.TRANSFORMER.INFERENCE_TH_TEST 0.37 \
  DATASETS.TEST '("inversetext_test",)'

The full set of thresholds used for the tables:

Backbone	Total-Text	Rot.Total-Text	Inverse-Text	CTW1500	ArT
ResNet-50	0.37	0.415	0.45	0.495	0.375
oCLIP	0.34	0.34	0.37	0.365	0.35
MonkeyOCRv2	0.405	0.395	0.37	0.375	0.355

tools/search_th.py sweeps INFERENCE_TH_TEST for a trained model and reports the best F-score:

python tools/search_th.py \
  --output-dir output/mkv2vitae_align/totaltext/direct_rotate \
  --test-dataset totaltext_poly_test --start 0.1 --end 0.5 --num-gpus 8

ICDAR19-ArT

ArT has no public test ground-truth. Evaluating an ArT config writes <OUTPUT_DIR>/inference/art_submit.json, which has to be uploaded to the ICDAR19-ArT evaluation server to obtain the P / R / F numbers reported above:

python tools/train_net.py --num-gpus 8 --eval-only \
  --config-file configs/DPText_DETR/ArT_Rotate/mkv2vitae_align.yaml \
  MODEL.WEIGHTS model_weight/dptext_mkv2vitae_art.pth

Acknowledgements

This project builds on DPText-DETR, AdelaiDet, detectron2, MMOCR, oCLIP, and MonkeyOCRv2.

License

The DPText-DETR / AdelaiDet sources this add-on patches are released for non-commercial use only (see LICENSE); the same restriction applies to this directory and to the released checkpoints.