search

aliyun / conditional-lane-detection

Apache License 2.0
396 stars 89 forks source link

Ground truth vs detected lane #72

Open NigussieAbate opened 8 months ago

NigussieAbate commented 8 months ago

Condlane-net Thank you author, for the great work! I am using CondlaneNet as a baseline framework to my research. I tried my own techniques to see the effect. But as you can see from the prediction result, the new result comes up with new lane which is not labeled in the ground truth (though it is clearly a lane line). Although I'm able to detect more lanes than the baseline, quantitative evaluation is not improved. Someone help me how I can justify it please?

dafodilrat-minuszero commented 8 months ago

I am having a hard time running the model and am facing a lot of issues with mmcv versions. Can you please walk me through how you got it set up.

haerrel commented 8 months ago

@NigussieAbate the problem that you describe is IIRC not related to #68. In my case the intermediate representation of the ground truth is wrong, therefore the model cannot learn these special cases.

What you are describing is that the dataset ground truth is not 100% accurate, what i also experienced during my research. You could update the ground truth for these samples

NigussieAbate commented 8 months ago

@haerrel thank you for your answers. I understood it is not related. I did that because I need you to replay on my questions. Would you please share your corrected GT in case you did corrections? Because updating the GT is not easy task for me. Thanks!

NigussieAbate commented 8 months ago

I am having a hard time running the model and am facing a lot of issues with mmcv versions. Can you please walk me through how you got it set up.

3

haerrel commented 8 months ago

@haerrel thank you for your answers. I understood it is not related. I did that because I need you to replay on my questions. Would you please share your corrected GT in case you did corrections? Because updating the GT is not easy task for me. Thanks!

I did not correct the ground truth data. I did corrections on the implementation, but only validated them on my custom dataset, that did not incorporate any of the given datasets.