SELMA DATASET

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Modalities (Sensors):

One of the major focuses of this dataset is the multimodal aspect of the data samples provided. In particular, we rendered 30’909 unique scenes in 27 environmental conditions, capturing the scene with the following sensors:

• 7 Color Cameras:
  • 3 frontal cameras (FRONT_LEFT, FRONT, FRONT_RIGHT), rectified and separated by 25cm. Can be used as 3 distinct stereo setups with baselines: 25cm (FL-F), 50cm (FL-FR) and 25cm (F-FR).
  • 1 desk camera (DESK).
  • 3 ring cameras (LEFT, RIGHT, BACK), facing in the remaining cardinal directions, providing a 360° view.
• 7 Depth Cameras, placed in coincidence with the color cameras to provide 2.5D information. The depth is provided in CARLA format, to get the depth in meters one must load the PNG file as an RGB image and perform the following computation: d = 1000*(R + G*256 + B*256*256)/(256*256*256 - 1).
• 3 LiDARS (64 channels, ~100k points per scan, provided in binary PLY format):
  • 1 top LiDARs (TOP), with a 360° view of the scene.
  • 2 front-lateral LiDARs placed on the headlights with 270° of aperture, overlapping in the front of the ego vehicle. The three LiDAR pointclouds can be merged to achieve higher density in such region.
• 1 Surface Normal Camera (estimated):
  • For a standard subset of our dataset (30909 samples in total) we also provide the pre-computed surface normals of the scene, as seen from the desk camera. They were estimated using the technique described in these articles [1, 2], starting from the ground truth depth.

Semantic Classes:

To better align with the common synthetic-to-real segmentation benchmark, Cityscapes, we generated the semantic labels for each sample (both for the cameras and the LiDARs) in a strict superset (36 classes in total) of the classes considered in the benchmark:

Class Name	Raw ID	Train ID	Cityscapes ID (35 classes)	Color (RGB)
Unlabeled	0	-1	0	■ (000, 000, 000)
Building	1	2	11	■ (070, 070, 070)
Fence	2	4	13	■ (190, 153, 153)
Other	3	-1	0	■ (000, 000, 000)
Pole	5	5	17	■ (153, 153, 153)
Road Line	6	0	7	■ (110, 190, 160)
Road	7	0	7	■ (128, 064, 128)
Sidewalk	8	1	8	■ (244, 035, 232)
Vegetation	9	8	21	■ (107, 142, 035)
Ego Vehicle	10	-1	1	■ (000, 000, 000)
Wall	11	3	12	■ (102, 102, 156)
Traffic Sign	12	7	20	■ (220, 220, 000)
Sky	13	10	23	■ (070, 130, 180)
Ground	14	-1	6	■ (081, 000, 081)
Bridge	15	-1	15	■ (150, 100, 100)
Rail Track	16	-1	10	■ (230, 150, 140)
Guard Rail	17	-1	14	■ (180, 165, 180)
Traffic Light	18	6	19	■ (250, 170, 030)
Static	19	-1	4	■ (110, 190, 160)
Dynamic	20	-1	5	■ (111, 074, 000)
Water	21	-1	36 (extra)	■ (045, 060, 150)
Terrain	22	9	22	■ (152, 251, 152)
Person	40	11	24	■ (220, 020, 060)
Rider	41	12	25	■ (255, 000, 000)
Car	100	13	26	■ (000, 000, 142)
Truck	101	14	27	■ (000, 000, 070)
Bus	102	15	28	■ (000, 060, 100)
Tram/Train	103	16	31	■ (000, 080, 100)
Motorcycle	104	17	32	■ (000, 000, 230)
Bicycle	105	18	33	■ (119, 011, 032)
Unknown	255	-1	0	■ (000, 000, 000)

Dataset Structure

The acquisitions are archived per scene and per sensor. A scene is a combination of a town and an environmental condition (weather and time-of-day). For instance, Town02_Opt_CloudyNoon is the scene acquired in Town02_Opt with a Cloudy sky at Noon. From the website, it is possible to select the data to download by combining different towns and environmental conditions. Furthermore, we offer the possibility to download only a subset of sensor data (e.g., only that acquired by the CAM_LEFT, i.e., the camera on the left side of the vehicle).

Weather	Clear	Cloudy	Wet	WetCloudy	SoftRain	MidRainy	HardRain
Time-of-day	Noon	Sunset	Night
Town	Town01_Opt	Town02_Opt	Town03_Opt	Town04_Opt	Town05_Opt	Town06_Opt	Town07_Opt	Town10HD_Opt

Download Instructions

1. Navigate to https://scanlab.dei.unipd.it/app/dataset

2. Login or register to access the data. We need you to register to monitor the traffic to/from our server and avoid overloading it. We will only use your e-mail for very limited service communications (new dataset releases, server status).

3. Select the subset of data you want to download using the selection tool available on the dataset page. We suggest starting with the SELMA random split (SELMA rand), that can be downloaded without setting any selection filter.

4. When prompted, download and unzip the archive. The zip contains the downloader toolkit: an executable downloader and a JSON file with the list of files to download.

5. Enable the execution of the downloader for your platform changing the permission accordingly (chmod +x downloader-<yourOS>.sh on Unix systems, double click and accept on Windows).

6. Execute the downloader and.. you’re set! You will find a new “CV” folder in the dataset-downloader-toolkit folder with all the data.

The data will be downloaded in a subfolder structure <SCENE_NAME>/<SENSOR_NAME> where <SCENE_NAME> is obtained as <Town_EnvironmentalCondition>.