search

cardwing / Codes-for-Lane-Detection

Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)
MIT License
1.04k stars 333 forks source link

evaluation script #105

Closed guozixunnicolas closed 5 years ago

guozixunnicolas commented 5 years ago

Hi,

May i know do you have the evaluation script in Python?

Could you please share it?

Best,

ZX

cardwing commented 5 years ago

After running test_lanenet.py, you get the probability maps from the model. To get the final performance, you just need to follow instructions in SCNN-Torch. More specifically, follow step 3 and 4 (i.e., get curve line from probability map and calculate precision, recall, and F-measure) in the testing phase.

guozixunnicolas commented 5 years ago

Thank you! it works now:)

harryhan618 commented 5 years ago

The evaluation for CULane is quite slow. It takes tens of minutes or even an hour to finish all the evaluation (9 different test list). I wanna check if this is normal.

cardwing commented 5 years ago

It is normal since there are 30k images in the testing set.

harryhan618 commented 5 years ago

@cardwing Thanks! Pardon me for inquiring the details. May I ask what parameters do you use for evaluation for CULane? Do you use the default setup that width_lane: 30 and iou_threshold: 0.5?

cardwing commented 5 years ago

Yes. You can just check the details in scripts of ENet-Label-Torch. The performance of the released model (F1-measure: 72.0) is higher than the reported value (F1-measure: 70.8).