CLIP4STR

This is a dedicated re-implementation of CLIP4STR: A Simple Baseline for Scene Text Recognition with Pre-trained Vision-Language Model .

Introduction
Installation
Results
Training
Inference
Citations
Acknowledgements

News

[02/05/2024] Add new CLIP4STR models pre-trained on DataComp-1B, LAION-2B, and DFN-5B. Add CLIP4STR models trained on RBU(6.5M).

Introduction

This is a third-party implementation of the paper CLIP4STR: A Simple Baseline for Scene Text Recognition with Pre-trained Vision-Language Model.

The framework of CLIP4STR. It has a visual branch and a cross-modal branch. The cross-modal branch refines the prediction of the visual branch for the final output. The text encoder is partially frozen.

CLIP4STR aims to build a scene text recognizer with the pre-trained vision-language model. In this re-implementation, we try to reproduce the performance of the original paper and evaluate the effectiveness of pre-trained VL models in the STR area.

Installation

Prepare data

First of all, you need to download the STR dataset.

We recommend you follow the instructions of PARSeq at its parseq/Datasets.md . The gdrive links are gdrive-link1 and gdrive-link2 from PARSeq.
For convenient, you can also download the STR dataset with real training images at BaiduYunPan str_dataset.
For the RBU(6.5M) training dataset, it is a combination of [the above STR dataset] + [val data of benchmarks (SVT, IIIT5K, IC13, IC15)] + [Union14M_L_lmdb_format]. For convenient, you can also download at BaiduYunPan str_dataset_ub.
weights of CLIP pre-trained models:
- CLIP-ViT-B/32
- CLIP-ViT-B/16
- CLIP-ViT-L/14
- OpenCLIP-ViT-B-16-DataComp-XL-s13B-b90K.bin
- OpenCLIP-ViT-L-14-DataComp-XL-s13B-b90K.bin
- OpenCLIP-ViT-H-14-laion2B-s32B-b79K.bin
- appleDFN5B-CLIP-ViT-H-14.bin
  - For models from huggingface.co, you should rename them as the shown names.

Generally, directories are organized as follows:

${ABSOLUTE_ROOT}
├── dataset
│   │
│   ├── str_dataset_ub
│   └── str_dataset           
│       ├── train
│       │   ├── real
│       │   └── synth
│       ├── val     
│       └── test
│
├── code              
│   │
│   └── clip4str
│
├── output (save the output of the program)
│
│
├── pretrained
│   └── clip (download the CLIP pre-trained weights and put them here)
│       └── ViT-B-16.pt
│
...

Dependency

Requires Python >= 3.8 and PyTorch >= 1.12. The following commands are tested on a Linux machine with CUDA Driver Version 525.105.17 and CUDA Version 11.3.

conda create --name clip4str python=3.8.5
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 -c pytorch
pip install -r requirements.txt

If you meet problems in continual training of an intermediate checkpoint, try to upgrade your PyTorch

conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia

Results

CLIP4STR pre-trained on OpenAI WIT-400M

CLIP4STR-B means using the CLIP-ViT-B/16 as the backbone, and CLIP4STR-L means using the CLIP-ViT-L/14 as the backbone.

Method	Train data	IIIT5K	SVT	IC13	IC15	IC15	SVTP	CUTE	HOST	WOST
		3,000	647	1,015	1,811	2,077	645	288	2,416	2,416
CLIP4STR-B	MJ+ST	97.70	95.36	96.06	87.47	84.02	91.47	94.44	80.01	86.75
CLIP4STR-L	MJ+ST	97.57	95.36	96.75	88.02	84.40	91.78	94.44	81.08	87.38
CLIP4STR-B	Real(3.3M)	99.20	98.30	98.23	91.44	90.61	96.90	99.65	77.36	87.87
CLIP4STR-L	Real(3.3M)	99.43	98.15	98.52	91.66	91.14	97.36	98.96	79.22	89.07

Method	Train data	COCO	ArT	Uber	Checkpoint
		9,825	35,149	80,551
CLIP4STR-B	MJ+ST	66.69	72.82	43.52	a5e3386222
CLIP4STR-L	MJ+ST	67.45	73.48	44.59	3544c362f0
CLIP4STR-B	Real(3.3M)	80.80	85.74	86.70	d70bde1f2d
CLIP4STR-L	Real(3.3M)	81.97	85.83	87.36	f125500adc

CLIP4STR pre-trained on DataComp-1B, LAION-2B, and DFN-5B

All models are trained on RBU(6.5M).

Method	Pre-train	Train	IIIT5K	SVT	IC13	IC15	IC15	SVTP	CUTE	HOST	WOST
			3,000	647	1,015	1,811	2,077	645	288	2,416	2,416
CLIP4STR-B	DC-1B	RBU	99.5	98.3	98.6	91.4	91.1	98.0	99.0	79.3	88.8
CLIP4STR-L	DC-1B	RBU	99.6	98.6	99.0	91.9	91.4	98.1	99.7	81.1	90.6
CLIP4STR-H	LAION-2B	RBU	99.7	98.6	98.9	91.6	91.1	98.5	99.7	80.6	90.0
CLIP4STR-H	DFN-5B	RBU	99.5	99.1	98.9	91.7	91.0	98.0	99.0	82.6	90.9

Method	Pre-train	Train	COCO	ArT	Uber	log	Checkpoint
			9,825	35,149	80,551
CLIP4STR-B	DC-1B	RBU	81.3	85.8	92.1	6e9fe947ac_log	6e9fe947ac, BaiduYun
CLIP4STR-L	DC-1B	RBU	82.7	86.4	92.2	3c9d881b88_log	3c9d881b88, BaiduYun
CLIP4STR-H	LAION-2B	RBU	82.5	86.2	91.2	5eef9f86e2_log	5eef9f86e2, BaiduYun
CLIP4STR-H	DFN-5B	RBU	83.0	86.4	91.7	3e942729b1_log	3e942729b1, BaiduYun

Training

Before training, you should set the path properly. Find the /PUT/YOUR/PATH/HERE in configs, scripts, strhub/vl_str, and strhub/str_adapter. For example, the /PUT/YOUR/PATH/HERE in the configs/main.yaml. Then replace it with your own path. A global searching and replacement is recommended.

For CLIP4STR with CLIP-ViT-B, refer to

bash scripts/vl4str_base.sh

8 NVIDIA GPUs with more than 24GB memory (per GPU) are required. For users with limited GPUs, you can change trainer.gpus=A, trainer.accumulate_grad_batches=B, and model.batch_size=C under the condition A * B * C = 1024 in the bash scripts. Do not modify the code, the PyTorch Lightning will handle the left.

For CLIP4STR with CLIP-ViT-L, refer to

bash scripts/vl4str_large.sh

We also provide the training script of CLIP4STR + Adapter described in the original paper,

bash scripts/str_adapter.sh

Inference

bash test.sh {gpu_id} {subpath_of_ckpt}

For example,

bash scripts/test.sh 0 clip4str_base16x16_d70bde1f2d.ckpt

If you want to read characters from an image, try:

bash test.sh {gpu_id} {subpath_of_ckpt} {image_folder_path}