"SA-DETR: Saliency Attention-based DETR for Salient Object Detection"
- PyTorch >=1.5.0
- Requirements
pip install -r requirements.txt
Fan, Deng-Ping, et al. "Salient objects in clutter." IEEE Transactions on Pattern Analysis and Machine Intelligence 45.2 (2022): 2344-2366.
python main_SOC.py \
--masks \
--no_aux_loss \
--output_dir "output_path" \
--epochs 200 \
--frozen_weights detr-r50-e632da11.pth (or --frozen_weights detr-r101-2c7b67e5.pth --backbone resnet101)
[--resume "output_checkpoint_path" --lr "lr" --lr_drop "lr_drop"]
python main_SOC.py --masks --no_aux_loss --eval
python pred_SOC.py --masks --no_aux_loss --eval
Perazzi, Federico, et al. "Saliency filters: Contrast based filtering for salient region detection." 2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012.
-
$N$ : Pixel numbers -
$Sal$ : Saliency(Output) map -
$G$ : GT map
Fan, Deng-Ping, et al. "Structure-measure: A new way to evaluate foreground maps." Proceedings of the IEEE international conference on computer vision. 2017.
-
$\alpha$ : Balanced parameter, [0, 1], (0.5 default) -
$S_o$ : Object-aware structural similarity -
$S_r$ : Region-aware structure similarity
Fan, Deng-Ping, et al. "Enhanced-alignment measure for binary foreground map evaluation." arXiv preprint arXiv:1805.10421 (2018).
-
$w, h$ : Width, height of map -
$\phi_{FM}$ : Enhanced alignment matrix of forground map
Ablation studies of Saliency Module(SM)
- Without SM, salient objects are not detected, or other objects are detected as salient.
- With SM, each attention map recognizes the shape of an object well, resulting in an accurate object-level mask.