ENet-Label-Torch is available now (a light-weight and effective lane detection model)

[cardwing]

Our ENet-Label-Torch has been released. More details can be found in my repo.

Key features:

(1) ENet-label is a light-weight lane detection model based on ENet and adopts self attention distillation (more details can be found in our paper which will be published soon).

(2) It has 20 × fewer parameters and runs 10 × faster compared to the state-of-the-art SCNN, and achieves 72.0 (F1-measure) on CULane testing set (better than SCNN which achieves 71.6).

(Do not hesitate to try our model!!!)

Performance on CULane testing set (F1-measure):

Category	SCNN-Torch	SCNN-Tensorflow	ENet-Label-Torch
Normal	90.6	90.2	90.7
Crowded	69.7	71.9	70.8
Night	66.1	64.6	65.9
No line	43.4	45.8	44.7
Shadow	66.9	73.8	70.6
Arrow	84.1	83.8	85.8
Dazzle light	58.5	59.5	64.4
Curve	64.4	63.4	65.4
Crossroad	1990	4137	2729
Total	71.6	71.3	72.0
Runtime(ms)	133.5	--	13.4
Parameter(M)	20.72	--	0.98

[ConerK]

和 LaneDetection_End2End 相比，哪个速度快精度高呢

[cardwing]

@ConerK Since SCNN is the best algorithm in TuSimple and ENet-Label-Torch outperforms SCNN in TuSimple, I think ENet-Label-Torch should outperform LaneDetection_End2End in terms of accuracy. As to speed, you can have a test.

[ConerK]

@cardwing OK，Thanks

[wvangansbeke]

@cardwing,

Thanks for letting me know. I will check it out.

Best, Wouter

[Msabih]

@cardwing

Could you please tell about the accuracy score that you obtained on tusimple and if you used only tusimple for training ?

[cardwing]

@Msabih ENet-label achieves 96.64% in TuSimple testing set, which outperforms SCNN (96.53%).

[Msabih]

@cardwing Thanks for the reply. Just another question. The tusimple benchmark provides y samples and lane points with respect to the image size of 720 x 1280. Did you train and test your model with image size of 720 x 1280 ?

If you resized the images to a lower resolution then you either have to resize the prediction images back to 720 x 1280 or you could also downsample the tusimple benchmark points and get evaluation ? Which is your approach or which one do you think is better ?

[cardwing]

I resize the input image to be 368 × 640, which follows the data processing of SCNN. However, a better solution is to remove the areas which do not have lanes, i.e., remove the upper areas of the image. And in this condition, you can use the full (cropped) image as input.