Computer Vision News - June 2016

Once the regression model was trained, the system builds on several steps to segment a given batch of images. First, the set of predefined segmentation is automatically generated for every image. Then, the regression prediction framework is applied to all resulting segmentations to estimate the quality of each result. Next, the system sorts all the images from highest to lowest predicted quality scores for the resulting segmentations. Finally, the system allocates the available human budget to create fine- Results The authors compared their work with: (1) Jain & Grauman (2013) whose system predicts how to best allocate a given budget of human time to annotate a batch of images; (2) segmentations collected from online layman crowd workers; (3) segmentations created by experts; (4) the proposed method itself, mixing human and computer efforts. Experiments demonstrate the advantage of relying on a mix approach over relying on either of the compared methods alone. Specifically, the proposed method eliminates the need for human annotation effort for an average of 44% of images (eliminating 30-60 minutes of human annotation time), while preserving the resulting segmentation quality achieved when relying exclusively on human input. It’s worth noting that the paper implements two systems that automatically decide, for a batch of images, when to replace 1) humans with computers, to create coarse segmentations required to initialize segmentation tools and 2) computers with humans, to create final, fine-grained segmentation. This is only a brief summary: if you want more details, the source code will soon be available here . Conclusions The authors’ experiment in all its task shows that their prediction module offers a sharp time saving in the human effort, while at the same time obtaining comparable quality levels of machine-driven algorithmic segmentation. Reduction in human intervention was therefore achieved without sacrificing the quality of the results. grained segmentations for the allotted number of images with the lowest predicted quality scores. Figure 2 above illustrates this process. Figure 2 Computer Vision News Research 19