The codes consist of three parts:
1 DDFM Inference: To generate the initial intermediate modality. This part of code comes from Zhao Z, Bai H, Zhu Y, et al. DDFM: denoising diffusion model for multi-modality image fusion[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 8082-8093. https://github.com/Zhaozixiang1228/MMIF-DDFM
2 Distillation stage: BMG pre-trained by distilling DDFM as its counterpart.
3 Fine-tuning stage: BMG tuned together with the feature extractor to suit the counting task and bridge two original modalities.
- Update on 25 May, 2025:
Share Link: https://pan.baidu.com/s/1pEW_77fOcbx6icZ3ndr8Pg?pwd=eysu
Containing:
- Density maps after fine-tuning on RGBT-CC;
- RGBT-CC dataset with initial broker modality and .npy files;
- DroneRGBT dataset with initial broker modality and .npy files;
- Broker modality images after fine-tuning on RGBT-CC;
- Weight files:
- Backbone pre-training weight (need loading in args in Fine-tuning/train.py);
- Broker Modality Generator (BMG) distillation weight from DDFM on RGBT-CC (need loading in args in Fine-tuning/models/bm.py);
- Broker Modality Generator (BMG) distillation weight from DDFM on DroneRGBT (need loading in args in Fine-tuning/models/bm.py);
- Model weight after fine-tuning on RGBT-CC;
- Model weight after fine-tuning on DroneRGBT.
Note: During fine-tuning, the format of .npy files is consistent with Bayesian Loss (https://github.com/ZhihengCV/Bayesian-Crowd-Counting): The first two columns represent the position of each annotation point. The third column records the distances between each annotation point and its nearest neighbors. The share link contains a script to generate .npy files from the original .npy files and the .npy files after processing.
Update on 22 Aug, 2025: Alternative link: https://drive.google.com/drive/folders/1SM-8bFW8ABG5Wlxv34dVw6PE0cm348ql?usp=sharing