Reproduce “DropClick: Semi-Automated One-Click Segmentation for Agricultural Robotic Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Patrick Zimmer, Michael Halstead, Chris McCool
DropClick, developed by researchers at the University of Bonn and CSIRO, is a semi-automated one-click segmentation system designed for agricultural robotic data, which predicts masks for both user-clicked and unclicked objects within a single network pass. It achieved comparable or superior mIoU to baseline methods with full clicks and maintained high performance even when approximately 50% of clicks were omitted, leading to substantial annotation effort savings.
Key insights
- A semi-automated system can effectively segment both clicked and unclicked objects concurrently, significantly reducing the required user input for segmentation tasks.
- High segmentation mask quality can be maintained even when a substantial percentage of input clicks are intentionally omitted (e.g., ~50% missing clicks), outperforming baseline methods like SAM2 in such scenarios.
- Pseudo-labels generated by DropClick with partial user input are robust enough to train downstream instance segmentation models in a semi-supervised manner, achieving performance nearly identical to models trained with all clicks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.