The B2B software market for size recommendation software consists of various players with different solutions. To determine the right size of clothing for customers, there are many different approaches. First there are picture- and app-based solutions. They often require downloads and only produce grey and generic body visualizations. Due to the limitation of only having one side view and one frontal view, they are prone to errors resulting from small changes in posture or clothing, which leads to low accuracy and robustness.
Benchmark: We decided to benchmark against the 2019 winner of the 3D Body Conference in Lugano which is the leading Computer Vision conference in our space. Almost every player on the market competes there.
Setup: For the benchmarking, the same input data was given to the Presize algorithm and the competitors algorithm.
Sample: 255 subjects with a mix of tight clothing, everyday clothing and demographics.
Ground truth: Body measurements from professional tailors who manually measured each subject according to the same ISO standard.
MAPE = Mean Absolute Percentage Error (deviation from ground truth in %)
Over all 255 subjects:
90% of subjects had a lower MAPE with Presize
55% lower MAPE with Presize
For the 99 subjects in tight clothing:
94% of subjects had a lower MAPE with Presize
67% lower MAPE with Presize
We combine up to 6 body measurements into one size recommendation. For example, if chest circumference is off 1 cm for a t-shirt, it’s very unlikely that all measurements (belly, shoulders, hips, arms…) are off and that they deviate in the same direction which would result in a wrong size recommendation.
Just like the rest of the competition, the competitor requires the user to take two pictures to predict body measurements. If you are looking for a solution that combines high convenience (recording a video is optional, good results in regular clothing and imperfect poses, no app download needed) and high accuracy, Presize is the right choice.