hciilab/derpn

A novel region proposal network for more general object detection ( including scene text detection ).

derpn's Issues

Code optimization

The paper is very interesting, and thank you very much for sharing the code.
I encountered loss nan when I used multi-batch training. After the check, code in generate_derpn_labels_targets_layer.py should be modified except for the assertion code.
Old
#############
for i in range(cls_num_ofchannel.shape[1]):
cls_weights[0,i, :, :].fill(cls_num_ofchannel[0,i])
for i in range(reg_num_ofchannel.shape[1]):
reg_weights[0,i, :, :].fill(reg_num_ofchannel[0,i])
#############
New
#############
for ith_im in range(num_images):
for i in range(cls_num_ofchannel.shape[1]):
cls_weights[ith_im,i, :, :].fill(cls_num_ofchannel[ith_im,i])
for i in range(reg_num_ofchannel.shape[1]):
reg_weights[ith_im,i, :, :].fill(reg_num_ofchannel[ith_im,i])
#############

Add code

If you wouldn't mind, please close this issue only after the code has been fully added. This issue is for tracking purposes.

about some code in tools/ron/generate_derpn_labels_targets_layer.py/self.eval_obs_region(gt_boxes)

def eval_obs_region(self, gt_boxes):
        height, width = self.height, self.width
        observe_region = np.zeros((height, width), np.int32)
        scaled_gt = gt_boxes[:,:4].copy()/float(self._feat_stride)#gt scaled to featrue map
        gt_w = scaled_gt[:, 2] -scaled_gt[:, 0]
        gt_h = scaled_gt[:, 3] -scaled_gt[:, 1]
        gt_cx = (scaled_gt[:, 2] + scaled_gt[:, 0])*0.5
        gt_cy = (scaled_gt[:, 3] + scaled_gt[:, 1])*0.5

        start_x = np.maximum((gt_cx - gt_w*self.extend_ratio*0.5).astype(np.int32), 0)
        end_x = np.minimum((gt_cx + gt_w*self.extend_ratio*0.5).astype(np.int32), width-1)
        start_y = np.maximum((gt_cy - gt_h*self.extend_ratio*0.5).astype(np.int32), 0)
        end_y = np.minimum((gt_cy + gt_h*self.extend_ratio*0.5).astype(np.int32), height-1)
        for ith in range(gt_boxes.shape[0]):                
            observe_region[start_y[ith]:end_y[ith]+1,start_x[ith]:end_x[ith]+1] = 1
        return observe_region, gt_cx, gt_cy`

`
The function of the above code, i think it is to reflect the gtbox into the feature map, and start_x,end_x ,start_y and end_y is to compute the coordinate in the feature map.
But what does the self.extend-ratio mean and why it is equal to 1.2?
And maybe i dismiss the real function of the above codes.
thanks a lo

Implement into Mask R-CNN

Hi, is it possible to implement DeRPN into Mask R-CNN?

The pre-trained model can not be download

Could you please fix the download link?
Can not move forward with the testing stage.

about eq (1) and eq(4)(5)

hello, i think eq (1) and eq(4)(5) are equivalent.

Just want the Region Proposal Network

Hey, I was wondering how to download your pre-trained Region Proposal Network. I don't want the classifier with it.