mxbh / mapformer Goto Github PK

I found your paper titled "MapFormer: Boosting Change Detection by Using Pre-change Information " very interesting. Could you please share experimental details and supplementary documents for this paper?

I wanted to reproduce the numbers, but I got very poor performance for the conditional BCD HRSCD dataset. Please find the log link and help me to reproduce the numbers.
https://drive.google.com/file/d/1PM9jO0wt0oc5JyjP_C-j-WmesmO_UmVY/view?usp=sharing

Here are the steps I performed.
I copied all the scripts from the tools folder to mapformer home directory.
Data preprocessing:
I noticed that for data, we need "-0M50-E080" added string only for tiles500 in the folder names of images, labels, and change. However, we don't need this for tiles2000
I run these lines
python convert_datasets/create_hrscd_tiles.py
--data_dir=./data/HRSCD/
--out_dir=./data/HRSCD/preprocessed/tiles500_2
--tile_size=500

Used for faster inference, internally splitted into 500px tiles within the model.
python convert_datasets/create_hrscd_tiles.py
--data_dir=./data/HRSCD/
--out_dir=./data/HRSCD/preprocessed/tiles2000_2
--tile_size=2000
Training: I could not download the weights provided at this link https://drive.google.com/drive/folders/1b7bwrInTW4VLEm27YawHOAMSMikga2Ia.
However, I downloaded the weights at this link https://download.openmmlab.com/mmsegmentation/v0.5/pretrain/segformer/mit_b2_20220624-66e8bf70.pth and renamed them to mit_b2.pth.
I modified the training script for the HRSCD dataset and ran this script.
python train.py ./configs/conditional_bcd/hrscd/mapformer.yaml
--work-dir=./runs/conditional_bcd/hrscd/mapformer

Inference: very poor performance.

Please find the attached log file here https://drive.google.com/file/d/1PM9jO0wt0oc5JyjP_C-j-WmesmO_UmVY/view?usp=sharing
Please look into it and help me to reproduce the numbers.

Regards:
Mustansar Fiaz