Update lane_detection_research.md

doc/research/paf24/perception/lane_detection_research.md

@@ -72,7 +72,9 @@ The best performing model on CurveLanes is **CondLSTR (ResNet-101)** [7]
 
 ### CondLSTR (ResNet-101)
 
-The paper "Generating Dynamic Kernels via Transformers for Lane Detection" presents a novel approach to lane detection, using dynamic convolutional kernels generated by transformers. This method achieves greater flexibility by adapting kernels to specific spatial features in images, which enhances detection accuracy in challenging environments. The dynamic kernel approach allows the model to perform well in varied lane geometries and lighting conditions. This framework demonstrates competitive performance on standard lane detection benchmarks, highlighting the potential of transformer-based architectures in dynamic feature extraction for computer vision tasks.
+The paper "Generating Dynamic Kernels via Transformers for Lane Detection" presents a novel approach to lane detection, using dynamic convolutional kernels generated by transformers.
+This method achieves greater flexibility by adapting kernels to specific spatial features in images, which enhances detection accuracy in challenging environments.
+The dynamic kernel approach allows the model to perform well in varied lane geometries and lighting conditions. This framework demonstrates competitive performance on standard lane detection benchmarks, highlighting the potential of transformer-based architectures in dynamic feature extraction for computer vision tasks.
 
 ![Image-CondLSTR](../../../assets/perception/CondLSTR(RESNet101)_Model_Overview.jpg)
 
@@ -98,6 +100,7 @@ For more details, view the paper [here](https://openaccess.thecvf.com//content/I
 - Improved Robustness: Due to the dynamic nature of the kernels, the model performs better under challenging conditions like occlusions and varying lane configurations, which often confuse traditional methods​
 
 **disadvantages:**
+
 - Computational Complexity: While the model offers significant performance improvements, the use of transformers and dynamic kernel generation may result in higher computational cost and slower inference times, especially in real-time applications​
 
 - Dependence on High-Quality Data: Like many deep learning models, this approach relies on large, well-labeled datasets for training. In environments where high-quality lane annotation is difficult or expensive to acquire, the model may struggle​
@@ -112,19 +115,18 @@ For more details, view the paper [here](https://openaccess.thecvf.com//content/I
 
 ![Comparison_on_dataset](/doc/assets/perception/Comparison_of_CurveLane_Dataset.jpg)
 
-It should be mentioned that the paper by Robin Karlsson et. al was also read, but it was not worth comparing it here as it was a selfsupervised learning model and was not trained on the above datasets but on a smaller dataset and used more sensors and cameras (6 cameras, 5 radars and 1 lidar) than allowed in the Qualify.
+It should be mentioned that the paper by Robin Karlsson et. al was also read, but it was not worth comparing it here as it was a selfsupervised learning model and was not trained on the above datasets but on a smaller dataset and used more sensors and cameras
+(6 cameras, 5 radars and 1 lidar) than allowed in the Qualify.
 But for interrested people you can find the paper [here](https://arxiv.org/pdf/2304.13242v2)
 
 ## Decision
 
 We chose to implement the CLRerNet-DLA34 and the CondLSTR(ResNet-101) because they have already implemented code bases to save time.
 
-
 [1]: https://paperswithcode.com/task/lane-detection#datasets
 [2]: https://paperswithcode.com/dataset/tusimple
 [3]: https://arxiv.org/pdf/2305.17271v2
 [4]: https://paperswithcode.com/dataset/culane
 [5]: https://arxiv.org/pdf/2305.08366v1
 [6]: https://paperswithcode.com/dataset/curvelanes
 [7]: https://openaccess.thecvf.com//content/ICCV2023/papers/Chen_Generating_Dynamic_Kernels_via_Transformers_for_Lane_Detection_ICCV_2023_paper.pdf
-