[TCSVT 2026] DiffRGBD: Diffusion-driven RGB-D Salient Object Detection with Temporal Modulation
🔗 IEEE Paper

• Python 3.10
• PyTorch 2.4.0

We provide the predicted saliency maps of:

• DiffRGBD (ours): Download Link (code: XHCL)
• Other RGB-D methods: Download Link (code: XHCL)

Evaluated on the following datasets:

• DUT
• LFSD
• NJU2K
• NLPR
• SIP
• SSD
• STERE1000

1. Download the pretrained backbone:

   • sam2_hiera_large.pt (code: XHCL)

2. Modify the checkpoint loading path in:./model/net.py

3. Set your dataset paths in the configuration file.

4. Run training: accelerate launch train.py
   --config config/camoDiffusion_352x352.yaml
   --num_epoch=YOUR_EPOCHS
   --batch_size=YOUR_BATCH_SIZE
   --gradient_accumulate_every=1

Download pretrained model: our_checkpoint(code: XHCL)

Run inference： accelerate launch sample.py
--config config/camoDiffusion_352x352.yaml
--results_folder YOUR_OUTPUT_PATH
--checkpoint YOUR_CHECKPOINT_PATH
--num_sample_steps 10
--target_dataset DATASET_NAME

We recommend using the following toolkit for evaluation: Evaluation Tool.

This work is built upon Camodiffusion and SAM2UNet. We sincerely thank the authors for their great contributions.

If you have any questions, encounter issues, or find bugs, please feel free to contact: shixiang_joy@163.com.