ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow

Paper | Code

ProtoFlow is a time-aware prototype dynamics framework for incremental remote sensing segmentation, designed to reduce catastrophic forgetting under class and domain shifts.

Installation

cd protoflow_reproduce
python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Preparing datasets

The project uses a generic image/mask folder format with split files:

data/processed/<dataset>/
  train.txt
  val.txt
  test.txt  # optional
  images/
    xxx.png
  masks/
    xxx.png

Each split file contains one sample per line:

images/xxx.png masks/xxx.png

Masks must contain integer class ids in [0, num_classes-1] and optional ignore id 255.

For large tiles, use:

python -m protoflow.tools.prepare_patches \
  --image-dir data/raw/deepglobe/images \
  --mask-dir data/raw/deepglobe/masks \
  --out-root data/processed/deepglobe \
  --split train \
  --patch-size 512 \
  --stride 512

Run the command for train/val/test as needed. For Vaihingen/Potsdam, use stride 256 as described in the paper.

Training commands

DeepGlobe class-incremental

python -m protoflow.tools.train --config configs/deepglobe.yaml

Vaihingen class-incremental

python -m protoflow.tools.train --config configs/vaihingen.yaml

Potsdam class-incremental

python -m protoflow.tools.train --config configs/potsdam.yaml

LoveDA domain-incremental

python -m protoflow.tools.train --config configs/loveda.yaml

Evaluation

python -m protoflow.tools.evaluate --config configs/deepglobe.yaml --checkpoint outputs/deepglobe_protoflow/checkpoints/final.pt

Analysis and plotting

The trainer writes JSON/CSV logs under outputs/<experiment>/metrics. Use:

python -m protoflow.tools.analyze_prototypes --run-dir outputs/deepglobe_protoflow
python -m protoflow.tools.plot_analysis --run-dir outputs/deepglobe_protoflow --out-dir outputs/deepglobe_protoflow/figures

Project layout

protoflow/
  configs/                         YAML configs for all paper protocols and smoke test
  protoflow/
    config.py                      YAML loading and config helpers
    datasets/                      generic folder dataset, incremental wrappers, transforms
    models/                        HRNet-W48, simple CNN, segmentation head
    engine/                        training, evaluation, losses, metrics, memory, prototype bank
    tools/                         CLI tools: train, eval, patchify, synthetic data, analysis plots
  scripts/run_smoke_test.sh         one-command smoke test
  requirements.txt

Citation

@misc{wu2026protoflowmitigatingforgettingclassincremental, 
      title={ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow}, 
      author={Jiekai Wu and Rong Fu and Chuangqi Li and Zijian Zhang and Guangxin Wu and Hao Zhang and Shiyin Lin and Jianyuan Ni and Yang Li and Dongxu Zhang and Amir H. Gandomi and Simon Fong and Pengbin Feng},
      year={2026},
      eprint={2604.03212},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.03212}, 
}