zeakey / iccv2019-fmeasure

Code accompanying the paper Optimizing the F-measure for Threshold-free Salient Object Detection.

Optimizing the F-measure for Threshold-free Salient Object Detection

Code accompanying the paper Optimizing the F-measure for Threshold-free Salient Object Detection.

Howto

1. Download and build caffe with python interface;
2. Download the MSRA-B dataset to data/ and the initial VGG weights to model/
3. Generate network and solver prototxt via python model/fdss.py;
4. 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.

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.

Example detection results

Several detection results. Our method results in high-contrast detections.

Stability against threshold

FLoss (solid lines) achieves high F-measure under a larger range of thresholds, presenting stability against the changing of threshold.

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.