Using Covariance Distribution to Improve Real-Time Lane Detection Segmentation Model
Real-time lane detection is very challenging due to very subtle and sparse supervisory lane signals in the image and lack of computational resources on the mobile devices. Current state-of-the-art semantic segmentation model has a large number of parameters, while small real-time memory footprint model still lack of accuracy and robustness on mobile device. To tack this problem, we introduce a covariance distribution module (CDM) to improve the accuracy of lightweight semantic segmentation network. The module optimize the covariance distribution of feature maps and it can be easily applied to most semantic models without modifying the structure of the model. Different from channel and space-based attention method, the proposed method has the same inference time with the original model, and it consumes very limited computing resources during model training. We use lightweight models such as ENet, CGNet, and ERFNet to verify CDM on two popular lane detection datasets (TuSimple, CULane). Notably, the CDM has proved to be effective in all lightweight models, and all results indicate that CDM can achieve obvious improvement in almost all real-time lane detection tasks
Year of publication: |
[2022]
|
---|---|
Authors: | Liu, Yian ; Xu, Ping ; Zhu, Lei ; Yan, Ming ; Xue, Linyun |
Publisher: |
[S.l.] : SSRN |
Saved in:
freely available
Saved in favorites
Similar items by person
-
Liu, Yian, (2011)
-
Zhu, Zhihong, (2019)
-
Xu, Ping, (2021)
- More ...