Thank you for sharing the approach. Could you please give an example or commit a demo on how you provide your datasets for training? Your train.py script used the DataLayer and reads data from data_dir='/home/wangyukang/dataset/' I think? But how is it formatted, what is the structure (image and gt in one folder? gt as images or as txt?)

CXX tools/caffe.cpp CXX tools/test_net.cpp CXX tools/finetune_net.cpp CXX examples/TextField/inference.cpp CXX .build_release/src/caffe/proto/caffe.pb.cc examples/TextField/inference.cpp: In function ‘int main(int, char**)’: examples/TextField/inference.cpp:43:3: error: reference to ‘shared_ptr’ is ambiguous shared_ptr <Net> net_; ^~~~~~~~~~

I managed to test your model which was already trained on total-text dataset, but I cannot reach your score. I want to know what is your environment? Such as GCC version, opencv-python version. Thanks!

源代码 top[0].data[...] = np.sum((self.distL1**2)*(self.weightPos + self.weightNeg)) / bottom[0].num / 2. / np.sum(self.weightPos + self.weightNeg) 当中bottom[0].num中的num是什么，好像Python当中没有这个

Hi, TextField is a great work, but I'm confused about the post processing:

We apply a simple dilation δ (with 3 × 3 structuring element) to group the representatives of the same text instance. This is followed by a connected component labeling that forms candidate text instances. The text superpixel grouping is depicted in line 17- 21 of Algorithm 1.

and I found these in your post processing code:

for (int row = 0; row < rows_; row++)
        {
            float* ending_p = ending.ptr<float>(row);
            float* parent_p = parent.ptr<float>(row);
            float* dict_p = dict.ptr<float>(row);
            for (int col = 0; col < cols_; col++)
            {
                if (ending_p[col] == 1)
                {
                    for (int dilDepth = 1; dilDepth <= min((int)(1*dict_p[2*col+1]-16), 12); dilDepth++)  //
                    {
                        p.x = row+(int)parent_p[2*col]*dilDepth;
                        p.y = col+(int)parent_p[2*col+1]*dilDepth;
                        if (p.x >= 0 && p.x <= rows_-1 && pt.y >= 0 && pt.y <= cols_-1)
                        {
                            float* merged_ending_p = merged_ending.ptr<float>(p.x);
                            merged_ending_p[p.y] = 1;
                        }
                    }
                }
            }
        }

I understand that the dilate element kernel_size = 3x3 is for grouping the representatives of text instance, and the above code is for grouping the rest children pixels belonging to this instance, but I don't understand this: for (int dilDepth = 1; dilDepth <= min((int)(1*dict_p[2*col+1]-16), 12); dilDepth++), and dilDepth stands for the deepest depth, but the magic numbers 16 and 12 represent for what meaning?

Thanks in advance for your reply!

I managed to retrain your approach with my own dataset and it performance quit well! However, the runtime/inference speed seems to be slower than compared with several other approaches (EAST e.g.) - especially if ported to a non GPU version. To you have any hints/ideas on how to improve the inference speed? Could the model retrained differently to better fit smaller inference scales?