On the Effectiveness of Visible Watermarks

Supplementary Material

 

We show input watermarked images and results for 50 images sampled randomly from each dataset (the total number of images in each dataset appears below). Note that the processing on each dataset used more images than just the sampled ones shows below, as detailed in the paper (Section 5).

* All stock images shown in our paper and supplementary material have been licensed from the corresponding stock services.

 

Results on Stock Images (Section 5.1)

Supplements Figure 6 in the paper. For each dataset we show 50 randomly sampled input (watermarked) images, and results for the first 20 of those images that we have licensed from the stock image providers. All the following results were produced completely automatically, without any user input.

 

 

"AdobeStock"

422 images. Source: http://stock.adobe.com/.

Sampled input images
Estimated watermark
Sampled watermark removal results

     
 

"123RF"

1376 images. Source: http://www.123rf.com/

Sampled input images
Estimated watermark
Sampled watermark removal results

     
 

"fotolia"

285 images. Source: http://www.fotolia.com/.

Sampled input images
Estimated watermark
Sampled watermark removal results

     
 

"CanStockPhoto"

4086 images. Source: http://www.canstockphoto.com/.

Sampled input images
Estimated watermark
Sampled watermark removal results

 

Results on Synthetic Datasets (Section 5.2):

Results on watermarked image collections we generated for the evaluation in the paper. Input and results are shown for 50 images sampled randomly from each dataset. In these examples the position of the watermark in each image was sampled randomly. We can then bootstrap the processing by getting from the user a rough bounding box around the watermark in one of the images ("user input" below). The rest of the processing is fully automatic.

 

 

"COCO_Copyright"

1000 images. Source: Microsoft COCO dataset (images), CanStockPhoto (watermark).

Sampled input images
User input
Estimated watermark
Sampled watermark removal results


     
 

"COCO_CVPR2017"

1000 images. Source: Microsoft COCO dataset (images), CVPR2017 (watermark).

Sampled input images
User input
Estimated watermark
Sampled watermark removal results

 

Robustness to Watermark Variations (Section 5.3)

The affect of different, per-image watermark variations on the removal attack (supplements Figure 5 in the paper).

COCO_CVPR2017 - 50 randomly sampled images from the COCO_CVPR2017 dataset.

COCO_Copyright - 50 randomly sampled images from the COCO_Copyright dataset.

Spatial perturbation - Example of the subtle spatial deformations we apply to the watermark to introduce geometric inconsistencies across the watermarked collection.