USODFuseNet

Model/Dataset Overview

Mode	Dataset Link	Pre-computed Saliency Maps	Model Weights	Pre-trained Weights
USOD	Train/Test	Saliency Maps	Weights	Pre-trained Weights
RGBD	Train/Test	Saliency Maps	Weights	Pre-trained Weights

Training

There are two settings: USOD (Underwater Salient Object Detection) and RGBD. Below, changing the training_scheme will enable either training.

The below command for USOD training -

python training.py \
    --lr 0.0005 \
    --epochs 26 \
    --f_name "USODFuseNet" \
    --n 4 \
    --b 16 \
    --sched 1 \
    --training_scheme "RGBD" \
    --salient_loss_weight 1.0 \
    --use_pretrained 1 \
    --checkpoint_name "SODAWideNet++" \ 
    --im_size 256

For RGBD training, use the following command -

python training.py \
    --lr 0.0005 \
    --epochs 21 \
    --f_name "USODFuseNetRGBD" \
    --n 4 \
    --b 16 \
    --sched 1 \
    --training_scheme "RGBD" \
    --salient_loss_weight 1.0 \
    --use_pretrained 1 \
    --checkpoint_name "SODAWideNet++" \ 
    --im_size 256

Inference

Download the trained weights and place them in a folder named checkpoints. Also, change checkpoint names in inference.py file (line numbers 24 & 25) for different modes.

1. Use the command below to produce a saliency map for a single image-depth pair. Remove the display flag from the above command to save the prediction.

python inference.py --mode single --input_path path/to/image.jpg --depth_path path/to/depth_image.jpg --display

2. Use the below command to run inference on a series of images in a folder and save the predicted saliency maps.

python inference.py --mode folder --input_path path/to/image_folder --depth_path path/to/depth_folder --output_dir path/to/output_folder