TRoM Benchmark
Welcome to the TRoM Vision Benchmark Suite!
AI & Self-driving Lab, Department of Computer Science and Technology, Tsinghua University, Beijing, China

Tsinghua Road Markings, called TRoM, is a dataset for recognition of road marking. This is the first dataset that focuses on road markings for self-driving or ADAS. Samples in the dataset were collected in Beijing municipality, China and the collection period lasted for over one month. It covers a diversity of traffic and weather conditions. In the current version of TRoM, 19 categories of the road markings were annotated for recognition use. This benchmark is aiming to develop amazing data-driven deep learning algorithms for self-driving & ADAS. The detailed information is listed below.


Xiaolong Liu, Zhidong Deng and Hongchao Lu

Some typical scenes in TRoM

Some typical scenes are listed here.

When using such dataset in your R&D work, we would like to be happy if you could cite the following article.
title={Benchmark for Road Marking Detection: Dataset Specification and Performance Baseline},   
author={Xiaolong Liu, Zhidong Deng, Hongchao Lu, Lele Cao},   


These datasets are made available for academic use only. We seriously take your privacy possibly involved in the datasets. If you find any problems, please kindly get contacted with us and we will then have removal of respective data from our server immediately.


   We are very grateful to all anonymous pedestrians for their support to this data collection.  This work was supported in part by the National Science Foundation of China (NSFC) under Grant Nos. 91420106, 90820305, and 60775040.


There are 19 categories indicated in the following table. The number of training set, validation set, and test set are 512, 100, 100, respectively.

Annotation toolkit

    The annotation toolkit contains a GUI developed via MATLAB. It is compatible with MATLAB versions installed on different operating systems. It reads raw images and allows users to conduct interactive through click-and-drag operations. Users can also browse annotated results or modify them if necessary. The annotation results are stored in an arbitrary directory.


On the basis of the TRoM data set released, we provide a few baselines, including FCN-16s, FCN-32s, FCN-8s, FCN-1s, ResNet, and our RPP model, to measure performance in road markings recognition.

  • 2016.05-2016.06 TRoM is collecting data!

  • 2017.07 Our Paper is accepted by ITSC!