search

MaybeShewill-CV / lanenet-lane-detection

Unofficial implemention of lanenet model for real time lane detection
Apache License 2.0
2.36k stars 886 forks source link

Testing on Different Images #508

Closed nbibin001 closed 3 years ago

nbibin001 commented 3 years ago

Hi, first of all thanks a lot for sharing this repository and detailed instructions for setting it up. It was very helpful to replicate and test it with the images in the data\tusimple_test_image folder. And with the default images present it worked just fine and I could replicate the results. But then I was trying to test it on some of the images from a different data set, it is not really working as expected. I am not seeing the detected lanes overlayed on the source image. I was just curious Is there something additional I need to do, or is it the limitation of the model, and I have to further train the model (transfer learning) to generalize it for my data set as well ? Really appreciate your help and thoughts ...

image

MaybeShewill-CV commented 3 years ago

@nbibin001 The postprocess script was only suitable for tusimple dataset you need to adjust it according to you own data:)

akshatbhutra commented 3 years ago

@MaybeShewill-CV can you please tell me how can I proceed with getting lane coordinates and then evaluating different performance metrics if I have only binary segmented output not instance segmented one. Thanks.

MaybeShewill-CV commented 3 years ago

@akshatbhutra You can only get all lane pixels via binary segmentation result. To obtain different lanes you need to do cluster yourself:)

akshatbhutra commented 3 years ago

@MaybeShewill-CV can you please give some reference for clustering? Thanks.

MaybeShewill-CV commented 3 years ago

@akshatbhutra You may google some traditional lane detection method to get some help:)

akshatbhutra commented 3 years ago

@MaybeShewill-CV is it now possible to print the summary of layers of LaneNet Model?

MaybeShewill-CV commented 3 years ago

@MaybeShewill-CV Tensorflow 2 was not supported here:)

MaybeShewill-CV commented 3 years ago

@nbibin001 Maybe try tensorboard :)

akshatbhutra commented 3 years ago

@MaybeShewill-CV Is there any reason why official tusimple evaluation script was not used? Thanks!

MaybeShewill-CV commented 3 years ago

@akshatbhutra I do not know the reason. Maybe you can get some help on tusimple's website:)

akshatbhutra commented 3 years ago

@MaybeShewill-CV can you please tell me what's the main difference in the architecture of the decoders of the two branches to perform two different tasks(binary segmentation and pixel embedding)? I could not understand. Thanks!

MaybeShewill-CV commented 3 years ago

@akshatbhutra One branch is for semantic segmentation the other if for instance segmentation:)

MaybeShewill-CV commented 3 years ago

@akshatbhutra You may look up cross entropy loss's api for help

akshatbhutra commented 3 years ago

Okay, thanks!