SenseEarth2020 - ChangeDetection

1st place in the Satellite Image Change Detection Challenge hosted by SenseTime.

Our Method

Task Description

Given two images of the same scene acquired at different times, we are required to mark the changed
and unchanged areas. Moreover, as for the changed areas, we need to annotate their detailed semantic masks.

The change detection task in this competition can be decomposed into two sub-tasks:

  • binary segmentation of changed and unchanged areas.
  • semantic segmentation of changed areas.



Pseudo Labeling

The core practice is using self-distillation strategy to assign pseudo labels to unchanged areas.

Specifically, in our experiments, predictions of five HRNet-based segmentation models are ensembled,
serving as pseudo labels of unchanged areas.

The overall training process can be summarized as:

  • Training multiple large segmentation models.
  • Ensembling their predictions on unchanged areas.
  • Training a smaller model with both labeled and pseudo labeled areas.

For more details, please refer to the
technical report
and presentation.

Getting Started


Description | Download [password: f3qq]

Pretrained Model

HRNet-W18 | HRNet-W40 | HRNet-W44 | HRNet-W48 | HRNet-W64

Final Trained Model

PSPNet-HRNet-W18 | PSPNet-HRNet-W40

File Organization

# store the whole dataset and pretrained backbones
mkdir -p data/dataset ; mkdir -p data/pretrained_models ;

# store the trained models
mkdir -p outdir/models ; 

# store the pseudo masks
mkdir -p outdir/masks/train/im1 ; mkdir -p outdir/masks/train/im2 ;

# store predictions of validation set and testing set
mkdir -p outdir/masks/val/im1 ; mkdir -p outdir/masks/val/im2 ;
mkdir -p outdir/masks/test/im1 ; mkdir -p outdir/masks/test/im2 ;

├── data
    ├── dataset                    # download from the link above
    │   ├── train                  # training set
    |   |   ├── im1
    |   |   └── ...
    │   └── val                    # the final testing set (without labels)
    └── pretrained_models
        ├── hrnet_w18.pth
        ├── hrnet_w40.pth
        └── ...


# Please refer to utils/ for more arguments
# If hardware supports, more backbones can be trained, such as hrnet_w44, hrnet_w48
CUDA_VISIBLE_DEVICES=0,1,2,3 python --backbone hrnet_w18 --pretrained --model pspnet --lightweight

Pseudo Labeling & Re-training

# This step is optional but important in performance improvement
# Modify the backbones, models and checkpoint paths in L20-40 in manually according to your saved models
# It is better to ensemble multiple trained models for pseudo labeling

# Pseudo labeling

# Re-training
CUDA_VISIBLE_DEVICES=0,1,2,3 python --backbone hrnet_w18 --pretrained --model pspnet --lightweight --use-pseudo-label


# Modify the backbones, models and checkpoint paths in L39-44 in manually according to your saved models
# Or simply use our final trained models