Placing Objects in Context via Inpainting for Out-of-distribution Segmentation

Codebase to realistically insert objects into images via diffusion inpainting models with applications to anomaly segmentation and dataset extension as presented in Placing Objects in Context via Inpainting for Out-of-distribution Segmentation.

---

Dependencies

This repository is built on top of different repositories for the generative pipeline and the anomaly segmentation analyses. Thus we recommend using different conda environments to install the necessary dependencies more easily.

POC pipeline

Our implementation of the Placing Objects in Context (POC) pipeline depends on two main blocks. An open vocabulary segmentor GSAM and an inpainting model based on StableDiffusion. based on StableDiffusion.

In order to run our pipeline, from the repo you may create a conda environment with the following commands:

# From within the cloned repo main folder
git clone https://github.com/naver/poc.git
cd poc/Grounded-Segment-Anything

export CUDA_HOME=/PATH/TO/CUDA_11.6
conda create -n POC python=3.10 -y
conda activate POC
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116
python -m pip install -e segment_anything
python -m pip install -e GroundingDINO
pip install diffusers[torch]==0.15.1

You also need to download the checkpoints for GSAM:

cd /MODEL/CHECKPOINT/FOLDER
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
wget https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

Then modify the corresponding paths in path_utils.py

Anomaly segmentation

To fine-tunine models with OOD data and perform anomaly segmentation please install the dependencies from RPL, Mask2Anomaly and RbA as detailed in their resepective documentation. Note that Mask2Anomaly and RbA can share the same environment given they are both built on top of Mask2Former.

Dataset extension

To train models with extended datasets we use the relis repository. Please install their dependencies to run mmsegmentation models.

---

Datasets

Existing benchmarks

To evaluate anomaly segmentation we take the evaluation sets provided in Mask2Anomaly via this link.

To fine-tune with Cityscapes and COCO, please download and prepare the datasets as described in Mask2Anomaly (all anomaly segmentation methods follow the same pre-processing).

For dataset extension, please download the PASCAL dataset (VOC2012) for evaluation.

POC Datasets

You can download the datasets generated in our work with the following link

wget https://download.europe.naverlabs.com/POC_Datasets/data.zip
unzip data.zip

The zip file does not contain the full images but only the local changes applied when inpainting an object. To obtain the final images you need to download the base datasets: Cityscapes or ACDC. Then you may run the poc/prepare_datasets.py script. To run it, first you need modify the root folder with the unzipped POC files and relevant base datasets. The POC datasets are released under the following license.

---

Getting started with POC

Download the SAM and GroundingDINO checkpoints for GSAM as well as the inpainting model:

cd /PATH/TO/MODELS/
# SAM weights
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
# GroundingDino
wget https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

## Inpainting model (optional to download locally)
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/stabilityai/stable-diffusion-2-inpainting

Fill the corresponding paths in poc/path_utils.py and run the poc/poc_pipeline_demo.ipynb or poc/test_poc_single_image.py.

---

Generating POC datasets

You may download Cityscapes, IDD and ACDC or you may use your own dataset.

To generate the datasets you may call the poc/generate_poc_dset.py. You may refer to the poc/launch_*.py scripts for usage examples to generate datasets with the same settings as used in our work.

You may also generate datasets with the Text2Image baselines using poc/generate_t2i_dset.py. In that case please download the stable diffusion base model

## Base text2image model
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/stabilityai/stable-diffusion-2-base

and add the path to poc/path_utils.py.

---

Fine-tuning and evaluating anomaly segmentation

Mask2Anomaly

Check the config files in Mask2Anomaly/configs/cityscapes/semantic-segmentation/ and replace the paths as appropriate. Then check the Mask2Anomaly/run.sh for the commands to fine-tune and evaluate models.

RPL

Modify the paths to datasets in RPL/rpl_corocl.code/config/config.py. Then check RPL/launch_fine-tune.sh and RPL/launch_test.sh

RbA

Check the config files of models in RbA/ckpts/ and commands to fine-tune and evaluate in RbA/launch_model_ft.sh and RPL/launch_test.sh respectively.

---

Learning new classes from POC data

To train models follow relis/mmsegmentation instructions. Once models are trained, fill the dataset paths, model checkpoints and configs in relis/path_dicts.py. In relis/launch_helpers.py you can find helper scripts to generate launch commands to evaluate models.

---

Reference

Placing Objects in Context via Inpainting for Out-of-distribution Segmentation (to appear at ECCV 2024)
Pau de Jorge, Riccardo Volpi, Puneet K. Dokania, Philip H. S. Torr and Grégory Rogez

    @InProceedings{deJorge_2024_ECCV,
    author = {de Jorge, Pau and Volpi, Riccardo and Dakania, Puneet K. and Torr, Philip H. S. and Gregory, Rogez},
    title = {Placing Objects in Context via Inpainting for Out-of-distribution Segmentation},
    booktitle = {The European Conference on Computer Vision (ECCV)},
    month = {October},
    year = {2024}
    }