lite deploy is throwing an exception "vector too long" in RunRecModel when using ch_PP-OCRv4_rec_infer

[prashant-saxena]

Hello,

Environment:
Windows 10
Visual Studio 17 2022
OpenCV 4.9.0 custom build
inference_lite_lib.win.x86.MSVC.C++_static.py37.full_publish.zip (downloaded)

The det is running fine with ch_PP-OCRv4_det_infer model.
The RunRecModel is either throwing an exception "vector too long" or terminating silently when using ch_PP-OCRv4_rec_infer.
The dictionary is ppocr_keys_v1.txt, test image is lite_demp.png and config.txt contents:

max_side_len  960
det_db_thresh  0.3
det_db_box_thresh  0.5
det_db_unclip_ratio  1.6
det_db_use_dilate 0
det_use_polygon_score 1
use_direction_classify  1
rec_image_height  48

I tried both the configurations MobileConfig and CxxConfig. det is working fine in both the cases but rec is throwing the same exception "vector too long"

[prashant-saxena]

The problem is this code in

PaddleOCR/deploy/lite/ocr_db_crnn.cc

Line 586 in 0525f6b

void rec(int argc, char **argv) {

int width = srcimg.cols;
int height = srcimg.rows;
std::vector<int> upper_left = {0, 0};
std::vector<int> upper_right = {width, 0};
std::vector<int> lower_right = {width, height};
std::vector<int> lower_left  = {0, height};
std::vector<std::vector<int>> box = {upper_left, upper_right, lower_right, lower_left};
std::vector<std::vector<std::vector<int>>> boxes = {box};

When passing lite_demo.png to process text, it's getting scaled down to a very small size by this call

PaddleOCR/deploy/lite/ocr_db_crnn.cc

Line 187 in 0525f6b

resize_img = CrnnResizeImg(crop_img, wh_ratio, rec_image_height);

I'm not sure about this behavior of passing an image as a single box containing multiple text blocks. I replaced the code by actually calculating the boxes first using

auto detpredictor = loadModel("data/ch_PP-OCRv4_det_infer", 2);
std::vector<double> times;
auto boxes = RunDetModel(detpredictor, srcimg, Config, &times);

This correctly calculates all the 63 text boxes that exist in the image.
The first box is correctly recognized

The detection visualized [ 0.96757 ]

When 2nd iteration of the loop executes

PaddleOCR/deploy/lite/ocr_db_crnn.cc

Line 178 in 0525f6b

for (int i = boxes.size() - 1; i >= 0; i--) {

control never reaches beyond this line and terminates the program

PaddleOCR/deploy/lite/ocr_db_crnn.cc

Line 192 in 0525f6b

std::unique_ptr<Tensor> input_tensor0(

I tried resetting the predictor at the end of the loop using

predictor_crnn.reset();

but to no avail. It seems the predictor is only working for one time. Any pointers?