OpenSketch: A Richly-Annotated Dataset of Product Design Sketches

Concept to presentation code, trained models and png images.

Download:

• Code repository (tensorflow/python)

  Contains the modified code of sketch2normal. In our paper we did not use user hints. It also contains python script to obtain oversketched lines from clean svg lines.

• sketch2normal_test_datasets.zip (157 MB)

  Contains test data:

  

          concept_all_lines_centered                          :concept sketch, all lines, sketch is centered
          concept_all_lines_original                          :concept sketch, all lines, sketch is as was originally positioned by an artists
          concept_presentation_ridges_silhouettes_centered    :concept and presentation sketches overlayed, rides and silhouette lines, sketch is centered
          concept_presentation_ridges_silhouettes_original    :concept and presentation sketches overlayed, rides and silhouette lines, sketch is as was originally positioned by an artists
          presentation_all_lines_centered                     :presentation sketch, all lines, sketch is centered
          presentation_all_lines_original                     :presentation sketch, all lines, sketch is as was originally positioned by an artists
          presentation_ridges_silhouettes_centered            :presentation sketch, rides and silhouette lines, sketch is centered
          presentation_ridges_silhouettes_original            :presentation sketch, rides and silhouette lines, sketch is as was originally positioned by an artists
          renderings                                          :silhoettes and ridges rendered from 3D models with camera parameters estimated with restricted camera model, sketch is centered
                  

  Each of this folders contains:

                      pngs_rendering_original : original width and opacity of strokes (width is scaled proportionally to the amge size)                          
                      pngs_rendering_uniform  : constant width and opacity of strokes
                      svgs_rendering_original : original width and opacity of strokes (width is scaled proportionally to the amge size)              
                      svgs_rendering_uniform  : constant width and opacity of strokes
                  

  Png images are obtained by rasterizing svg files with cairosvg.

• sketch2normal_predictions.zip (1.22 GB)

  Contains predictions on test data:

                      concept_all_lines_centered                          
                      concept_presentation_ridges_silhouettes_centered    
                      presentation_all_lines_centered                     
                      presentation_ridges_silhouettes_centered            
                      renderings                                          
                  

  Each of this folders contains:

                      pngs_rendering_original : 
                          D1,D2,D3,D4 predictions
                      pngs_rendering_uniform  : 
                          D1,D2,D3,D4 predictions
                  

• checkpoints_sketch2normal.zip (122 GB)

  Contains trained models on D1,D2,D3,D4.