Web companies selling customized garments like jackets, shirts and pants need to inspect products before shipment to customers. They want to compare the actual value of different longitudinal measures with those entered by customers when placing the command. This is called preshiment sizing or inspection, and is a common practice in the apparel industry. Current practice is workers sample products and take a dozen of measurements, since it is a time consuming process. Moreover, there’s also variability depending on the person.
We have developed a machine vision system that automatically obtains accurate measures from several views of each type of garment. Some of then can not be actually measured because of garment self occlusions, but are estimated from the value of other measures by means of a regressor trained with the grountruth of a number of samples. Results show an average mean absolute error well below 1 cm in all measures.
What’s inside ?
- garment segmentation and key point detection
- learned regressor to correct deviations from image to groundtruth measures
- regression of unseen measures from known ones
Joan Serrat, Felipe Lumbreras, Pau Li