Zero-Shot Scene Change Detection

AAAI 2025

Kyusik Cho Dong Yeop Kim Euntai Kim
Yonsei University

TL;DR: We present a novel, training-free approach to scene change detection by leveraging a tracking model.

Prerequisites

Pretrained Weights

Download the DEVA weights from this repository.
Download the SAM weights from this repository.
Place the downloaded weights in the model_weights folder.

Dataset Preparation

Download the datasets and set the file structure as follows.
Only the test dataset is required.
Edit dataset/path_config.py to set the dataset directory path.

ChangeSim

ChangSim
├── Query
│   └── Query_Seq_Test
│       ├── Warehouse_6
│       │   ├── Seq_0
│       │   │   └── ...
│       │   └── Seq_1
│       │       └── ...
│       ├──  ...
│       │
│       └── Warehouse_8
│           ├── Seq_0
│           │   └── ...
│           └── Seq_1
│               └── ...
│
└── Ref
    └── Ref_Seq_Test
        ├── Warehouse_6
        │   ├── Seq_0
        │   │   └── ...
        │   └── Seq_1
        │       └── ...
        ├──  ...
        │
        └── Warehouse_8
            ├── Seq_0
            │   └── ...
            └── Seq_1
                └── ...

VL-CMU-CD

VL-CMU-CD-binary255
└── test
    ├── t0
    ├── t1
    └── mask

PCD
├── TSUNAMI
│   ├── t0
│   ├── t1
│   ├── ground_truth
│   └── mask
└── GSV
    ├── t0
    ├── t1
    ├── ground_truth
    └── mask

Evaluation

ChangeSim

$ python main.py
$ python main.py --changesim_subset dark
$ python main.py --changesim_subset dust

VL_CMU_CD

$ python main.py --dataset VL_CMU_CD

$ python main.py --dataset TSUNAMI
$ python main.py --dataset GSV

Acknowledgements

Our pytorch implementation is based on the following projects and repos.