Awesome
<h1 align="center">Optimizing the F-measure for Threshold-free Salient Object Detection</h1>Code accompanying the paper Optimizing the F-measure for Threshold-free Salient Object Detection.
<div width=280px align="center"> <a href="http://kaizhao.net/fmeasure"> <img src="http://data.kaizhao.net/projects/fmeasure-saliency/qr-code.png" width=150px> </a> <a href="http://data.kaizhao.net/publications/iccv2019fmeasure.pdf"> <img src="http://data.kaizhao.net/projects/fmeasure-saliency/paper-thumbnail.png" width=130px> </a> </div>Howto
- Download and build caffe with python interface;
- Download the MSRA-B dataset to
data/and the initial VGG weights tomodel/ - Generate network and solver prototxt via
python model/fdss.py; - Start training the DSS+FLoss model with
python train.py --solver tmp/fdss_beta0.80_aug_solver.pt
Loss surface
The proposed FLoss holds considerable gradients even in the saturated area, resulting in polarized predictions that are stable against the threshold.
<p align="center"> <img src="http://data.kaizhao.net/projects/fmeasure-saliency/loss-surface.svg" width=100%> </p> <p align="center"> Loss surface of FLoss (left), Log-FLoss (mid) and Cross-entropy loss (right). FLoss holds larger gradients in the saturated area, leading to high-contrast predictions. </p>Example detection results
<p align="center"> <img src="http://data.kaizhao.net/projects/fmeasure-saliency/example-detections.png" width=800px> </p> <p align="center"> Several detection results. Our method results in high-contrast detections. </p>Stability against threshold
<p align="center"> <img src="http://data.kaizhao.net/projects/fmeasure-saliency/f-thres.svg" width=400px> </p> <p align="center"> FLoss (solid lines) achieves high F-measure under a larger range of thresholds, presenting stability against the changing of threshold. </p>Pretrained models
For pretrained models and evaluation results, please visit http://kaizhao.net/fmeasure.
If you have any problem using this code, please contact Kai Zhao.