ALIFE

This is the implementation of the paper "ALIFE: Adaptive Logit Regularizer and Feature Replay for Incremental Semantic Segmentation".

For detailed information, please check out our paper [arXiv].

Requirements

Python >= 3.6
PyTorch >= 1.3.0
yacs (https://github.com/rbgirshick/yacs)

Getting started

git clone https://github.com/cvlab-yonsei/ALIFE.git
cd ALIFE

mkdir checkpoints logs
mkdir -p datasets/voc/19-1-ov datasets/voc/15-5-ov datasets/voc/15-5s-ov
mkdir -p datasets/ade/100-50-ov datasets/ade/50-ov datasets/ade/100-10-ov

Datasets

The structure should be organized as follows:

├─ ALIFE
└─ data
    ├─ ADEChallengeData2016
    └─ VOCdevkit

Training

Example commands

bash scripts/voc/alife.sh   # RUN ALIFE   on PASCAL VOC (overlapped setting)
bash scripts/ade/alife-m.sh # RUN ALIFE-M on ADE20K     (overlapped setting)

NOTE: we also provide individual scripts for each step of ALIFE (e.g., base.sh, step1.sh, step2.sh, and step3.sh). You may need to modify those scripts for running a specific scenario (Please see ./scripts/). You can change config options by modifying the corresponding config files (./configs/) or typing 'KEY VALUE' pairs after '--opts' on the command line within the scripts.

Acknowledgements

Our codes are partly based on the following repositories.