Evaluation tools for segmentation and edge/centerline detection.
-
Segmentation Metrics:
- Global Accuracy (G)
- Class Average Accuracy (C)
- Mean IOU (I/U)
-
Sensitivity and Specificity Metrics:
- Precision (P)
- Recall (R)
- F-score (F)
Released results:
Note: The PyTorch implementation with the same loss achieves lower performances than the Caffe implementation. So, we suggest to set the loss mode as
focal
in the configuration filetrain_deepcrack.sh
.
Outputs | bT | G | C | I/U | P | R | F |
---|---|---|---|---|---|---|---|
DeepCrack-BN | 0.31 | 0.9873 | 0.9196 | 0.8643 | 0.8582 | 0.8456 | 0.8518 |
DeepCrack-GF | 0.48 | 0.9888 | 0.9261 | 0.8778 | 0.8795 | 0.8575 | 0.8684 |
Side-output 1 | 0.43 | 0.9836 | 0.8930 | 0.8298 | 0.8208 | 0.7939 | 0.8071 |
Side-output 2 | 0.42 | 0.9863 | 0.9093 | 0.8543 | 0.8537 | 0.8250 | 0.8391 |
Side-output 3 | 0.36 | 0.9854 | 0.9110 | 0.8482 | 0.8334 | 0.8295 | 0.8315 |
Side-output 4 | 0.36 | 0.9823 | 0.8989 | 0.8228 | 0.7886 | 0.8077 | 0.7980 |
Side-output 5 | 0.38 | 0.9735 | 0.8814 | 0.7663 | 0.6646 | 0.7807 | 0.7180 |
For comparisons, you can download our predicted images and evaluation files from google drive:
deepcrack
|__ evaluation
| |__ ...
|__ test_latest
|__images
|__ ...
*_image.png
: input images,*_label_viz.png
: ground truth,*_fused.png
: outputs of fused layer,*_gf.png
: refined predictions by guided filter, see the codetools/guided_filter.py
,*_side1.png
: side output 1,*_side2.png
: side output 2,*_side3.png
: side output 3,*_side4.png
: side output 4,*_side5.png
: side output 5,
TODO: CRF refinement module will be released soon...