About testing result

[Endless-Hao]

hello, i test the results use the prerain-model you offered, but i find the result is very worse .In some clear lane mark picture also not good,i want to know why.And is this model your last testing model? And 是否存在坐标的转换问题，我获得的测试结果，利用SCNN的tools生成结果图的。下面是我其中一张。

[Endless-Hao]

@cardwing Look the pic is ok ,but in the real world picture is not good.

[cardwing]

@Endless-Hao, what is the size of your input image? And do you test the pre-trained model and get the final F1-measure (71.3)? It seems that the probability map does not match the original image.

[cardwing]

I use the pre-trained model and get satisfactory prediction results (see the following figures). I guess there is something wrong with your testing script.

[Endless-Hao]

@cardwing input pic is CULane data size 1640590 ,and get result is from SCNN-tensorflow is 800288. And then i put them ,use the SCNN-original tools. Get the results like the picture i put in the web.

[Endless-Hao]

@cardwing Do you mean i should change the test input size to from 590 1640 to 288 800?

[cardwing]

@Endless-Hao, you should first check if the F1-measure of the pre-trained model is 71.3. The original input image should be ok.

[Endless-Hao]

@cardwing hello! I use SCNN original evalution get these: file: ./output/vgg_SCNN_DULR_w9_iou0.5.txt tp: 0 fp: 111702 fn: 0 precision: 0 recall: -1 Fmeasure: 0 All the step i follow your step.Also i change the input to 288 800,i get the result still not good.i don't know why.

[cardwing]

@Endless-Hao, can you share the folder of probability maps which is generated by test_lanenet.py? It is obvious that there is something wrong with the detection results.

[Endless-Hao]

The folder are like these: predicts_SCNN_test_final vgg_SCNN_DULR_w9 home pch test driver_37_30frame driver_100_30frame driver_193_90frame 06042010_0511.MP4 00000.exist.txt 00000_1_avg.png 00000_2_avg.png 00000_3_avg.png 00000_4_avg.png

[cardwing]

@Endless-Hao, can you attach the original folder here? I need the whole file to check which part is wrong. You can put the link which stores the generated probability maps here.

[Endless-Hao]

@cardwing hello you are a good 老哥 Thanks.This is my original folder: problem.zip

[cardwing]

@Endless-Hao, I have found that the generated probability maps are different these of mine. You can either test my generated probability maps or re-train the model and have a test then. It is not easy to find out the bug only using few sample probability maps.

[Endless-Hao]

@cardwing i look all the results, also i change the input ,but get different result which are all not good. Do you konw is it roughly what caused it? I use the data and model the same as you. If there is not solution at last ,i think i must retrain it. I just feel strange about this. The steps are same as you.

[cardwing]

@Endless-Hao, I have downloaded the pre-trained model and regenerate the probability maps. The results are still good and the F1-measure is exactly the same as what is listed in the table in README. I recommend that you should check the evaluation process as well as the probability map generation process.

[Endless-Hao]

@cardwing i have another problem, can you debug. I want to see the every step. Your code in my computer ,can not debug. When i debug, it just say"pyder debugger: process 26490 is connecting" only this sentence.

[Endless-Hao]

@cardwing i also find, when i test it. If there is no line in the picture, the exist.txt also 1 1 1 1. This is also strange thing.

[Endless-Hao]

@cardwing i use you result you upload. I find that your results is very good. The last generate picture is very good. the results you upload actually good.

[Endless-Hao]

@cardwing but i can not get these. Do you have any little change in your code you forget it ?

[cardwing]

@Endless-Hao, everything is ok in my local server, including debugging and running (I download the codes and perform testing). I also check the uploaded codes and find nothing is wrong. The main difference between your generated probability maps and mine is the position of detected lanes. You may have to re-train the model and check the performance then.

[Endless-Hao]

@cardwing i also find a strange condition. Every time i test the pictures, the same model and the same code will get the different results. The differences are a little big.

[Endless-Hao]

@cardwing i test all for three times, get three different results for them.Like blow:

[cardwing]

@Endless-Hao, it is weird. You should check the code carefully since it is impossible to get different outputs using the same input and same model.

[Endless-Hao]

@cardwing this is what i am so strange. If this can not solve, i think i will also get not good results to use the model that i train my own . I continue to check the code. Thinks. If i get the new condition, i would talk with you. Thank you for your help.

[cardwing]

^_^

[KevinWalkerFan]

@Endless-Hao can u share the test code that u use to test the visual performance just on the pic?

[BoSeal]

i also find, when i test it. If there is no line in the picture, the exist.txt also 1 1 1 1. This is also strange thing. @Endless-Hao ,Hello, I have the same problem as you. How did you solve this problem?