A python implementation of edge eval.
The logic of the code is almost the same as that of the origin MATLAB implementation (see References). The core code is used directly edge-eval-python.
I just added some optimizations for ease of use. Currently, BSDS, NYUD, BIPED, UDED can be used in one click, and other datasets can be easily eval by specifying GT.
- Python3
- Numpy
- Scipy >= 1.6.0
- g++
- Matplotlib
git clone https://github.com/Walstruzz/edge_eval_python.git
cd edge_eval_pythonMost of the code in this folder is copied from davidstutz/extended-berkeley-segmentation-benchmark.
Actually, there is a more efficient function in Scipy that can solve the CSA problem without compiling the following cxx codes...
cd cxx/src
source build.shsave your result like this: https://github.com/Li-yachuan/CTFN-pytorch-master/blob/main/test.py.
Make sure that there are two folders mat and png in the results folder to save the two forms of the prediction results
Let's say this folder is /result
Then the folder structure should be:
\result
....|-png
....|...|-a.png
....|...|-b.png
....|....:
....|-mat
........|-a.mat
........|-b.mat
:
Download the project and unzip the GT in the folder .\GT. They are available here
####3.1 The simplest eval just use
python eval.py [input dir] -d [dataset name] -ffor example:
python eval.py \result -d BSDS -fyou can eval the result of BSDS500 in dir \result
For higher automation, we implement automatic detection of folders with simple cycle. We use the T and limit parameters to control the sleep time and total wait time for the folder traversal, the default of which is 0.5 hours and 8 hours.
####3.2 No constant monitoring
If you only need to eval existing results, you can use notwait to cancel the automatic check and keep waiting, as follows:
python eval.py \result -d BSDS -nw -f####3.3 Default dataset
If you can include the dataset name in the path under eval, then you don't have to specify the dataset individually:
python eval.py \result\bsds-result -nw -f####3.4 multiple paths
You can specify multiple paths to eval multiple sets of results at the same time.You need to place multiple paths inside the "", separated by Spaces. There is no support for evaling multiple datasets at the same time, so multiple results should be on the same dataset.
python eval.py "\result\bsds-result \result\bsds-result2" -nw -f####3.5 light version & full version
Evaluation of edge detection is known to be a very time-consuming process, in extreme cases even slower than training, such as on densely labeled datasets like BIPED.
To speed up eval, we divided eval into two versions, light and full, by controlling the sampling frequency of the threshold. The full version is currently the most commonly used thrs=99, and thrs=9 in light.
light version:
python eval.py "\result\bsds-result \result\bsds-result2" -nwfull version:
python eval.py "\result\bsds-result \result\bsds-result2" -nw -fThe light version is faster, but the accuracy will be about 0.5% lower than the full version, but the relative accuracy will not change, so we can use the light version to pick the best result, and then use the full version to get its true accuracy.
####3.6 show result
To see the results better, use the following command.
for full version:
python show.py "\result\bsds-result" -ffor light version:
python show.py "\result\bsds-result"- The edges of the image are 1 and the background is 0. black background.
Note (same as edge-eval-python.)
- Because of the difference in calculation precision and the sensitivity of NMS threshold, the edge images may be slightly different.
match_edge_mapssamples points randomly (so as Matlab).- Python and Matlab index files in different order, resulting in different order of
eval_bdry_img.txt. - Python version is slower than Matlab version. Should I implement more functions in
cxx/lib/solve_cas.so?