1. Tusimple dataset
Features: Located on the expressway, the weather is fine, the lane lines are clear, and the lane lines are marked with dots. (ground_truth:json format) (Provides large-scale image data with instance-level lane annotations. But it is not suitable for video instance lane detection.)
1. Lane lines are actually not just markings on the road, dotted lines are treated as a A solid line is used for processing. Information such as double solid lines, white lines, and yellow lines is not marked here.
2. Each line is actually a set of coordinates of a point sequence, not an area set. The
main collection area is on foreign highways (installed near the license plate of the truck), the image is not distorted,
moderate weather conditions, different times of the day - 2 lanes / 3 lanes / 4 lanes / or more - different traffic conditions, each picture provides the first 19 frames of the picture (but these 19 frames are unmarked) lane lines are not classified, (the shooting angle should be outside the car ), GT is marked in json format in dot form. 20 frames are extracted in 1s. The shooting angle direction is very close to the driving direction of the car
Image size: 1280x720, number of images: 6408 (number: train: 3626; test: 2782)
download address: https://github.com/TuSimple/tusimple-benchmark/issues/3 (official), https://github. com/TuSimple/tusimple-benchmark/tree/master/doc/lane_detection
Baidu network disk link: https://pan.baidu.com/s/1i1IUxYI48tP5y-p2t-sVOA?pwd=1wy0 extraction code: 1wy0 (unofficial)
sample Processing: https://blog.csdn.net/qingfengxd1/article/details/108738651
Evaluation indicators: https://github.com/TuSimple/tusimple-benchmark/blob/master/evaluate/lane.py
2. CULane dataset
Features: Including crowded, night, wireless, shadow and other eight kinds of difficult detection situations, marking up to 4 lane lines. The data set is collected by cameras installed in six different cars operated by different drivers, and the road is located in Beijing. The dataset collects more than 55 hours of video data and extracts 133,235 frames of images.
The data sample is shown in Figure 1 below. Among them, the test set is divided into ordinary and the remaining 8 challenging types (the scene is different).
For situations where lane markings are obscured or invisible by vehicles, we still annotate the lanes contextually, and we also want the algorithm to be able to distinguish obstacles on the road. (Provide large-scale image data with instance-level lane annotations. However, it is not suitable for video instance lane detection.)
Beijing urban roads and expressway shooting locations, the front camera in the car has no obvious distortion correction traces, and the lane marked is the current lane And the left and right lanes (at most), do not distinguish between dashed and solid lines, and mark them in the form of dots (txt).
Picture size: 1640x590, number of pictures: 133235 (88880 images are divided into training set, 9675 images are divided into verification set, and 34680 images are divided into test set.) Download address:
https://xingangpan.github.io/projects /CULane.html (official)
sample processing: https://blog.csdn.net/weixin_43726913/article/details/115330065
evaluation indicators: https://github.com/ZJULearning/resa
3. BDD100k dataset
Features: A very comprehensive dataset from the University of Berkeley that also includes lane lines. Contains day and night weather conditions of 4 districts in the United States, 2D 8-category lane lines. (Annotations for lane instances are not provided—multiple lanes of the same type are not separated and annotated with a single label on each frame.) Curb, Crosswalk, Double White, Double Yellow, Double Other Color, Single White, Single Yellow, single other colors 8 main categories. These videos were collected from different places in the United States. As shown in the figure above, the database covers different weather conditions, including sunny, cloudy and rainy days, and at different times of day and night. Car front camera, what you see is what you get, city roads and highways.
Recently, the Berkeley AI Lab published the largest and most diverse open source video dataset in the CV field – the BDD100K dataset. The dataset consists of 100,000 videos, each video is about 40 seconds, 720P, 30fps, and the total time is more than 1,100 hours. The BDD100K data set extracts key frames at the 10th second of each video and marks them.
Lane markings are important road instructions for drivers. They are also key cues for driving direction and positioning for autonomous driving systems when GPS or maps do not have precise global coverage. The dataset classifies lane markings into two types based on their indication of vehicles within the lane. Vertical lane markings (marked in red in the image below) indicate markings in the direction of travel along their lane. Parallel lane markings (marked in blue in the image below) indicate signs for vehicles to stop within the lane. We also provide attributes for markers, such as solid vs. dashed, double vs. single. Annotation mask at the pixel level in label form
Picture size: 1280x720, number of pictures: 10w
Download address: https://bdd-data.berkeley.edu/ (official)
Link: https://pan.baidu.com/s/1ue7RuAitqKUTSAAHNPpkFg Extraction code: cdiy (unofficial)
Link to the mini version of BDD100K: https://pan.baidu.com/s/1r7FbMX2_OcDKsfrhujSfAQ Extraction code: yc90
Popular models in the current dataset: https://github.com/SysCV/bdd100k-models/
Data processing: https:/ /github.com/cardwing/Codes-for-Lane-Detection/tree/master/ENet-BDD100K-Torch
paper: BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning
Evaluation Indicators:
4. ApolloScape dataset
Features: This large-scale dataset contains a diverse set of stereoscopic video sequences recorded in street views from different cities with high-quality pixel-level annotations of more than 110 000 frames. (Annotation of lane instances is not provided—on each frame, multiple lanes of the same type are not separated and annotated with a single label.) WYSIWYG, pixel-level annotation mask contains 35 lane marking categories (
optional merged), among which there are 10 types of lanes, mainly domestic urban roads, and the camera is located on the vehicle body. Purpose: HD map construction or update process.
Picture size: 3384x2710, number of pictures: 200k (113653)
download address: http://apolloscape.auto/lane_segmentation.html (official), https://github.com/ApolloScapeAuto/dataset-api/blob/master/lane_segmentation/ LanemarkDiscription.pdf (introduction)
evaluation metrics: http://apolloscape.auto/lane_segmentation.html#to_evaluation_href
5. CurveLanes dataset
Features: Huawei curve detection data set, training set 100,000, verification set 20,000, test level 30,000 (the test set does not provide labels). The road is highly curved, and the labels are in dotted json format. Compared with BDD, Tusimple, and culane have more lanes. Lane classes are not distinguished. The camera is taken by the front camera inside the car.
Picture size: 2650x1440, number of pictures: 15w
Download address: https://github.com/SoulmateB/CurveLanes
Data processing: https://github.com/pandamax/Parse_Curvelanes
Evaluation indicators: To be provided.
6. LLAMAS data set
Features: 14 highway records were used, each about 25 km long
Sensor data was collected to label images in our dataset. foreign high speed. Bosch's lane marking data set, the labeling is not manual labeling, but generated by high-precision maps, all of which are high-speed scenes. Annotated over 100 meters, the camera is positioned roughly at the license plate on the outside of the car. The color image additionally provides the grayscale image of the original image, a total of 4 categories, the label is in point json format (three-dimensional coordinates), the left virtual lane line, the right virtual lane line, the left solid lane line, and the right solid lane line
Image size: 1276 x 717, number of images: 79113 (training set: 58269 validation set: 20844 test set: 20929 (no labels)) download
address: https://unsupervised-llamas.com/llamas/login/?next=/ llamas/download (registration required)
Evaluation metrics: https://github.com/karstenBehrendt/unsupervised_llamas/tree/master/evaluation
7. VPGNet Dataset
Features: Contains daytime (non-rainy, rainy, heavy rainy) and night data, as well as various types of lane markings, three-week driving, and other different types of lane markings (left-turn arrows, straight arrows, zebra crossings, etc.). 21097 road images in Seoul, South Korea, labels and original image format: mat, a total of 17 categories (8 lane categories, 9 road marking categories) manually labeled lanes and road corners. The corner points are connected to form a polygon forming pixel-level mask annotations for each object. The dataset was captured under severe weather conditions with a camera installed inside the car (middle). Urban traffic scene filmed in central Seoul, camera may capture windshield wipers
Picture size: 1288×728, number of pictures: 20k
Download address: If you would like to download the VPGNet dataset, please fill out a survey. We will send you an e-mail with a download link. (You need to fill out the questionnaire)
data Processing: https://blog.csdn.net/weixin_42840360/article/details/116995504
Paper address: VPGNet: Vanishing Point Guided Network for Lane and Road Marking Detection and Recognition
model Address: https://github.com/SeokjuLee/VPGNet
Evaluation metrics: https://github.com/SeokjuLee/VPGNet
8. VIL-100 Dataset
Features: It contains 100 videos with 10,000 frames of images, covering 10 classic scenes such as crowded road conditions, broken lane lines, smog, and night. We downsample the frame rate of all videos from the original 30fps to 10fps, providing instance-level annotations for all lane lines in each frame. The tags json and mask are both provided, considering the collection of 10 typical scenes in the data: normal, crowded, detour, damaged road, shadow, road marking, glare, haze, night, intersection. Of the 100 videos, 97 were captured by a single front-facing camera mounted near the rearview mirror. The remaining 3 videos are collected from the internet. There are 10 categories of lane lines in total, (single white solid line, single white dashed line, single yellow solid line, single yellow dashed line, double white solid line, double yellow solid line, double yellow dashed line, double white solid dashed line, double white dashed line solid line, double white dashed line Solid white and yellow.) single white solid, single white dotted, single yellow solid, single yellow dotted, double white solid, double yellow solid, double yellow dotted, double white solid dotted, double white dotted solid, double white solid w For each video, place a series of anchor points along the centerline of each lane in each frame and store
them in a json file. Points along each lane are stored in a set, which provides instance-level annotations. We then fit each set of points to a curve
through a third-order polynomial, and extend it to lane areas with a certain width. As a rule of thumb, on a 1,920 × 1,080 frame, choose a width of 30 pixels. For low-resolution
frames, the width is reduced proportionally.
Image size: 1920 × 1080, number of images: 10k
Download address: https://pan.baidu.com/s/1hFPKt4az6AiMmsV4c9Odmg - yyl7 (official)
paper address: VIL-100: A New Dataset and A Baseline Model for Video Instance Lane Detection
model address: https://github.com/yujun0-0/MMA-Net
evaluation index: https://github.com/yujun0-0/MMA-Net/tree/main/evaluation
9. Jiqing Expressway dataset
Features: This dataset is a multi-lane detection dataset that can be used to test and evaluate multi-lane detection algorithms. There are 40 video clips in the dataset, each video clip lasts 3 minutes, the frame rate is 30 fps, road images with different lighting intensities and different road conditions (upstream, downhill, tunnel, culvert, ramp, etc.). The driving recorder is similar to the shooting angle, high-speed domestic, for shadowed or vehicle-obstructed lanes, there are no markings or only partial markings (only a few cases). In addition, the dataset only annotates the lane lines in the direction of road traffic and does not include the reverse lane. Also, lane annotations for ramp sections may not be perfect. The lanes are not divided into categories, the labels are marked in txt format, and the original image video
Picture size: 1920×1080, number of pictures: 210610
Download address: https://github.com/vonsj0210/Multi-Lane-Detection-Dataset-with-Ground-Truth
Evaluation indicator: https://github.com/vonsj0210/ Multi-Lane-Detection-Dataset-with-Ground-Truth
10. Ceymo dataset
Features: Road markings belonging to 11 categories are manually annotated as polygons using the labelme annotation tool. Each image has a JSON file containing the coordinates of the polygons containing the road markings present
in that image. In addition to polygon annotation JSON format, we also provide bounding box annotation XML format and pixel-level segmentation mask PNG format to facilitate different road marking detection methods. Filming abroad; driving recorder
Image size: 1920 × 1080, number of images: 2887, train set: 2099, test set: 788 images
download address: https://github.com/oshadajay/CeyMo
paper address: CeyMo: See More on Roads – A Novel Benchmark Dataset for Road Marking Detection
evaluation indicators: https://github.com/oshadajay/CeyMo
11. SDLane DatasetDataset
Features: The SDLane dataset is a novel lane marking dataset for autonomous driving. We provide high-resolution images of 1920 x 1208 pixels, capturing challenging scenarios of highways and urban areas. The dataset consists of 39K training images and 4K testing images with accurate ground truth labels. For each scene, we manually annotate the 2D lane geometry of all visible lane markings on the road. Furthermore, to better infer the position of ego vehicles, we annotate the index of each lane relative to the leftmost lane marker. Exterior camera, South Korea. Label json format, point form, lane line regardless of category
Picture size: 1920 x 1208, number of pictures: 43k
Download address: https://www.42dot.ai/akit/dataset
Evaluation index: https://www.42dot.ai/akit/dataset
Comparison summary: Among them, TuSimple is challenging Low, the scene is mostly a highway, new papers like to use it to verify the feasibility. In contrast, the scenes of CULane are complex, many of which are located in the urban area of Beijing, and the difficulty is relatively high. At present, in the latest academic papers I have seen, most of them use TuSimple and CULane data sets for performance comparison, and some articles use LLAMAS.
Evaluation Index
When judging True or False, there are two main ways:
end point, by judging whether the distance between the endpoints of the line and its enclosing area exceeds the threshold
IOU, directly calculate the overlapping area of the IOU
