Scene-Text-Removal

Scene-Text-Removal copied to clipboard

@neouyghur You can try to get the mask from the background image and the input image.

Nov 20 '19 09:11 HCIILAB

When I use train_img_7634.png and try to get the diff, using simple L2 norm, the result is like this (from left to right, input (w/ text), target (w/o text), difference (black if there's difference))

It seems pixels values non-text regions are often changed dramatically between input/target. Since computing accurate masks should have high impact on calculating loss functions, could you also provide the masks that you have used in your codes?

Nov 21 '19 03:11 naoto0804

@naoto0804 That is the reason why I ask the mask.

Nov 21 '19 08:11 neouyghur

@neouyghur Great! I wish you could have shown the result;

Nov 21 '19 08:11 naoto0804

@naoto0804 I used the following code.

masks_refine_gt = np.greater(np.mean(np.abs(img.astype(np.float32) - img_gt.astype(np.float32)), axis=-1), self.mask_threshold).astype(np.uint8) # Threshold is set to 25

Nov 27 '19 06:11 neouyghur

@neouyghur You can try to get the mask from the background image and the input image.

@HCIILAB @zhangshuaitao

Although we can generate the mask by input and group truth difference, however for text detection evaluation we still need the ground truth bounding box. You also reported the related results. Without the official ground truth bounding box, it is hard to finish the complete comparison. Please provide a ground truth bounding box. Thanks.

Dec 03 '19 06:12 neouyghur

@neouyghur You can try to get the mask from the background image and the input image.

@HCIILAB @zhangshuaitao

Although we can generate the mask by input and group truth difference, however for text detection evaluation we still need the ground truth bounding box. You also reported the related results. Without the official ground truth bounding box, it is hard to finish the complete comparison. Please provide a ground truth bounding box. Thanks.

Have got the masks and ground truth bounding box... Please share with me as well

@neouyghur @naoto0804 @neouyghur

Apr 04 '23 04:04 UsamaRehman001