This is the official PyTorch implementation for SupeRGB-D: Zero-shot Instance Segmentation in Cluttered Indoor Environments, IEEE RAL paper.

In this work, we explore zero-shot instance segmentation (ZSIS) from RGB-D data to identify unseen objects in a semantic category-agnostic manner. We introduce a zero-shot split for Tabletop Objects Dataset (TOD-Z) to enable this study and present a method that uses annotated objects to learn the ``objectness'' of pixels and generalize to unseen object categories in cluttered indoor environments.

Our method, SupeRGB-D, groups pixels into small patches based on geometric cues and learns to merge the patches in a deep agglomerative clustering fashion. An overview of our method is illustrated here:

1. Python environment using env.yml.

    git clone https://github.com/evinpinar/supergb-d.git
    cd supergb-d
    conda env create --file env.yml
    conda activate supergbd

2. Install the TOD dataset from original repo and the TOD-Z ids from here. Preprocess the data to extract superpixels, training features and generate the ground truth. Fix the datapaths according to your local configuration.

    python data/preprocess_data_full.py # set up the number of threads according to your cpu
    # optionally, you can also only run data/process.py for single thread. 

3. Train the merger network.

    python src/model_train.py --cfg configs/run_local.yaml

4. Test the trained model.

    python src/model_eval.py --cfg configs/run_local.yaml

An example checkpoint is provided here which is based on 128 super-pixels and trained without DINO features.

If you find this code helpful, please consider citing:

@ARTICLE{ornek23,
  author={{\"O}rnek, Evin P{\i}nar and Krishnan, Aravindhan K and Gayaka, Shreekant and Kuo, Cheng-Hao and Sen, Arnie and Navab, Nassir and Tombari, Federico},
  journal={IEEE Robotics and Automation Letters}, 
  title={SupeRGB-D: Zero-Shot Instance Segmentation in Cluttered Indoor Environments}, 
  year={2023},
  volume={8},
  number={6},
  pages={3709-3716},
  doi={10.1109/LRA.2023.3271527}}

This repository contains code parts that are based on UOIS-Net and Davis-2017. We thank the authors for making their code available.