Multi-Object Discovery by Low-Dimensional Object Motion (ICCV 2023)

Project Page | Paper

This is the code for our ICCV 2023 paper:

Multi-Object Discovery by Low-Dimensional Object Motion
Sadra Safadoust and Fatma Güney

Requirements

1. Create the conda environment and install the requirements:

conda create -n mos python=3.8
conda activate mos
conda install -y pytorch=1.12.1 torchvision=0.13.1 cudatoolkit=11.3 -c pytorch
conda install -y kornia jupyter tensorboard timm einops scikit-learn scikit-image openexr-python tqdm -c conda-forge
conda install -y gcc_linux-64=7 gxx_linux-64=7 fontconfig matplotlib
pip install cvbase opencv-python filelock

2. Install Mask2Former:

python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
cd src/mask2former/modeling/pixel_decoder/ops
sh make.sh

If you face any problems while installing Mask2Former, please refer to the installation steps in the Mask2Former and Detectron2 repositories.

Datasets

CLEVR/ClevrTex and MOVi

Please follow the steps here for the CLEVR, ClevrTex, and MOVi datasets.

DAVIS2017

Download and extract the DAVIS2017 dataset into src/data/DAVIS2017. Download the motion annotations from here and extract them into /src/data/DAVIS2017/Annotations_unsupervised_motion. Use motion-grouping repository to create the flows for this dataset.

KITTI

Download the KITTI raw dataset into src/data/KITTI/KITTI-Raw. Calculate the flows using RAFT into src/data/KITTI/RAFT_FLOWS. Note that we use png flows for KITTI. Download segmentation labels from this repository.

---

The data directory structure should look like this:

├── src
    ├── data
        ├── movi_a
            ├── train
            ├── validation
        ├── movi_c
            ├── train
            ├── validation
        ├── movi_d
            ├── train
            ├── validation
        ├── movi_d
            ├── train
            ├── validation
        ├── moving_clevr
            ├── tar
                ├── CLEVRMOV_clevr_v2_*.tar
                ...
        ├── moving_clevrtex
            ├── tar
                ├── CLEVRMOV_full_old_ten_slow_short_*.tar
                ...
        ├── DAVIS2017
            ├── JPEGImages
            ├── Annotations_unsupervised_motion
            ├── Flows_gap4
            ...
        ├── KITTI
            ├── KITTI-Raw
                ├── 2011_09_26
                ...
            ├── RAFT_FLOWS
                ├── 2011_09_26
                ...
            ├── KITTI_test

Training

You can train the model on the datasets by running the corresponding scripts. E.g. for movi_c, run ./scripts/train_movi_c.sh

Inference

You can evaluate the model on the synthetic datasets by running the corresponding scripts. E.g. for movi_c, run ./scripts/eval_movi_c.sh and set the model weights path in the script to the path of the segmentation model.

Trained Models

You can download the trained models from here.

Acknowledgements

The structure of this code is largely based on probable-motion repository. Many thanks to the authors for sharing their code.