Improving the Estimation of Object mass from images

Image2mass and Pix2Vox++ are works that focus on the estimation of mass and the shape of objects in images, respectively. Having the aid of the Pix2Vox++ in mind, the proposed work aims to further explore the image2mass’s architecture by modifying its mass estimation pipeline. This pipeline contains a Geometry Module which estimates a thickness mask and 14 geometric features and uses them to predict the density and volume of objects, which result in the mass value when multiplied.Firstly, in this work we present a method to turn the output of the Pix2Vox++ into a thickness mask and 14 features, analogous to image2mass. Secondly, it is fine-tuned the Pix2Vox++ architecture with different configurations, using a dataset called GraspNet. The three best Pix2Vox++ models are then chosen to substitute, or aid, the Geometry Module, and the image2mass is fine-tuned using the same dataset. Additionally, texture randomization is executed to evaluate the generalization of results given the variability of the input data.The results demonstrate that the Pix2Vox++ help the image2mass to achieve a better performance. The fine-tuning configurations of both models, however, should be further explored.