| Summary: | Semantic mapping has advanced greatly since its inception as a research field, to now being able to identify poses and segment objects. As an extension of semantic segmentation problem, there is still an unexplored field of identifying joints of an articulated object within an image. In this project, our main contributions are to re-train a semantic segmentation network on a smaller subset of items which can be considered prismatic or revolute. With a DeepLabv3-Inception network with a ResNet101 backbone, we report best pixelwise accuracy of 0.931 and mIOU of 0.606. while training on 2 object classes from the ADE20K dataset. This preliminary result shows the viability of such an approach, and future work might entail exploring different loss functions; different neural network architecture and expanding the definition to encompass more items from the ADE20K dataset.
|
|---|