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

bisenetv2训练的远处结果太粗怎么办？ #440

Closed pauperonway closed 4 years ago

pauperonway commented 4 years ago

@MaybeShewill-CV bisenetv2训练后预测结果如下图：(好像有波纹一样的边缘) image 想要改进一下，有什么建议吗？

pythonmjs commented 4 years ago

@pauperonway Have you trained model on your own dataset?

pauperonway commented 4 years ago

@pauperonway Have you trained model on your own dataset?

可以修改config里的yaml，使用generate_tusimple_dataset.py生成自己的数据集，如果有问题，可以debug一下代码。

MaybeShewill-CV commented 4 years ago

@pauperonway 你可以尝试训练的时候crop比较大的patch 而不是采用[512, 256].

pauperonway commented 4 years ago

@MaybeShewill-CV 我现在的训练样本是1080P。。。 yaml: 10 TRAIN_CROP_SIZE: [1920, 1080] # crop size for training 11 EVAL_CROP_SIZE: [1920, 1080] # crop size for evaluating

pauperonway commented 4 years ago

@MaybeShewill-CV bisenetv2 的detail branch理论上应该对小目标比较友好啊，我回家再想想，哈哈哈

MaybeShewill-CV commented 4 years ago

@pauperonway 锯齿主要是由于下采样造成的 你预测的时候可以采用图像分块预测 不要直接resize可能效果会好一点

pauperonway commented 4 years ago

@MaybeShewill-CV 之前预测为了提高帧率，把图像resize到720，锯齿明显，改为原图1080P大小，如下图（效果正常） image

MaybeShewill-CV commented 4 years ago

@pauperonway 赞！

suyali commented 4 years ago

@pauperonway tusimple数据的标定转换的结果不是对虚车道线进行了连线的处理吗？ 为什么你检测输出的线条仍然是虚线？

pauperonway commented 3 years ago

@suyali 我使用的自己的数据集。