README.md

TERDNet

Transformer Encoder–Recurrent Decoder Network for Scene Change Detection ICRA 2026 Official Implementation

Overview

This repository provides the official implementation of TERDNet, proposed in:

TERDNet: Transformer Encoder–Recurrent Decoder Network for Scene Change Detection (ICRA 2026)

Architecture

TERDNet consists of:

• Transformer-based encoder
• Correlation-aware feature fusion module
• 3-gate GRU recurrent decoder
• Progressive upsampling module

Datasets

We follow the official splits of the following Scene Change Detection benchmarks:

• VL-CMU-CD
  http://vl-cmu-cd.cs.cmu.edu/

• TSUNAMI
  https://github.com/SakuradaK/ChangeDetectionDataset

• PSCD
  https://github.com/KeiSakurada/PSCD

• ChangeSim
  https://github.com/AI-ChangeSim/ChangeSim

Pretrained Models

This implementation supports pretrained transformer backbones.

• Segment Anything Model (SAM) https://github.com/facebookresearch/segment-anything

Training

Example training command:

python -u src/train.py \
    --train-dataset VL_CMU_CD \
    --test-dataset VL_CMU_CD \
    --data-cv 0 \
    --input-size 1024 \
    --model vit_b_terdnet \
    --fusion-type ours \
    --pyramid-levels 4 \
    --decoder-iters 3 \
    --warmup \
    --loss-weight

Acknowledgement

This implementation is built upon:

• C-3PO https://github.com/wgcban/C-3PO

• Segment Anything Model (SAM) https://github.com/facebookresearch/segment-anything

We thank the authors for releasing their code and pretrained models.

Citation

@InProceedings{Yoon_2026_ICRA,
    author    = {Yoon, Jiae and Kim, Ue-Hwan},
    title     = {TERDNet: Transformer Encoder-Recurrent Decoder Network for Scene Change Detection},
    booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
    month     = {June},
    year      = {2026},
}