LIRMM-BL 3D urban object scan dataset

Dataset Description :

This data set is part of a thesis project on detection, localization and monitoring of urban objects in 3D LIDAR scan taken continuously at ground level. The main goal is to be able to take a point cloud corresponding to a georeferenced scan of a city and precisely localize all the urban objects such as traffic signs, cars, trees, etc… and collect informations about their status, like a tree’s growth or a damaged light pole.
The datas consists of 3d point clouds of urban objects divided in 8 classes :

  1. Trees
  2. Cars
  3. Traffic signs/lights
  4. Poles
  5. Persons
  6. Building
  7. Noise/Scan artifact
First we searched for public annotated datasets with occurrences of object from those classes. We found the following datasets :

Links :

Sydney urban dataset :
Kevin Lai dataset :
Paris rue Madame dataset :

For each of these datasets, we extracted the examples, ie point clouds, who could fit under one of the class we previously defined. We then regrouped all of them in one base that we used as a training dataset.

LIDAR acquisition description :

Second we performed an urban LIDAR acquisition with the help of the Leica Pegasus backpack. The scanned consist of a small residential area, a tramway station and trees. We manually isolated 160 urban objects from the scan with the following classes repartition :

  1. 75 trees
  2. 39 cars
  3. 8 traffic signs/lights
  4. 23 poles
  5. 15 persons

In our experiments, we used this dataset as testing set and we use the network PointNet to classify them. The results are reported in the following table. More details about the experiments can be found in our article.


Made by Gaetan Fine 2018

Contact :