pytorch-dense-correspondence

Why we included this project

This is the reference implementation for Dense Object Nets, a self-supervised method that learns dense visual descriptors for objects a robot has never seen, including deformable ones. Instead of training a separate model for each grasp or task, the network builds a consistent per-pixel descriptor space, so a robot can locate and grasp a specific point on an object even after the object has been bent or rearranged. The repo includes a step-by-step tutorial, a Docker workflow, and pre-trained models, and training a descriptor network takes roughly 20 minutes, so you can go from a clean environment to a working model in about half an hour. It is a research codebase rather than a polished product, so expect Jupyter notebooks and a documented data-processing pipeline instead of a clean API. For teams exploring task-agnostic object representations for manipulation, the training and evaluation scripts are a solid place to start.