{}
Transfer learning using Hybrid Semantic Change Detection Data
We provide the weights used in the experiments of our CVPR'25 paper The Change You Want To Detect.
Dual U-Net
The model is a relatively simple Dual U-Net composed of two nearly identical parallel U-Nets. One responsible for semantic segmentation, the other for binary change detection. Besides using skip connections in each U-Net, extracted features from the "semantic encoder" are also transmitted to the "change detection decoder".
Both images are sequentially and independently passed through the "semantic U-Net", that produce a semantic map for each image, and extracted features are stored. Then images are concatenated, passed through the "change detection U-Net" while injecting the previously stored features and a binary change map is produced.
The backbones are ResNet-50 pretrained on ImageNet.
Model checkpoint
Here we provide the weights for our Dual U-Net that have been obtained after a pre-training on our Hybrid Semantic Change Dataset FSC-180k. They can be used for fine-tuning or inference on real change dataset using our code and specifically passing the .ckpt file as --run_id variable.