Last Modified: Octuber 21, 2019

jzolfaghari@cvc.uab.es
gvillalonga@cvc.uab.es

http://synthia-dataset.net

SYNTHIA: The SYNTHetic collection of Imagery and Annotations

This package contains the SYNTHIA_AL subset. This is a video stream generated at 25 FPS.

The number of classes here presented covers those defined in the table below, extending the classes

Instance segmentation groundtruth, 2D bounding boxes, 3D bounding boxes and depth information are provided!

Please, if you use this data for research purposes, consider citing our ICCVW 2019 paper:

@article{bengar2019temporal,
  title={Temporal Coherence for Active Learning in Videos},
  author={Zolfaghari Bengar, Javad and Gonzalez-Garcia, Abel and Villalonga, Gabriel and Raducanu, 
  Bogdan and Aghdam, Hamed H and Mozerov, Mikhail and Lopez, Antonio M and van de Weijer, Joost},
  journal={arXiv preprint arXiv:1908.11757},
  year={2019}
}


DESCRIPTION:


The package contains the following data,


	labels_kitti

		folder containing txt files (one per image). The content of each file is:
		<object class> <NULL> <NULL> <observation angle of object> <2D bounding box in the format of [xmin, ymin, xmax, ymax]> <3D bounding box in the format of [xpos, ypos, zpos, width, height, length, angle]>

	SemSeg
		folder containing png files (one per image). Annotations are given in three channels. First
		channel contains the class of that pixel (see the table below). The other two channels contain
		the unique ID of the instance for those objects that are dynamic (cars, pedestrians, etc.).

	RGB
		folder containing standard 640x480 RGB images 


	modified_labels_kitti
		
		folder containing txt files (one per image). Only the annotation for images in which the objects are appeared with
		more than 50 pixels. The content of each file is:
		<object class> <NULL> <NULL> <observation angle of object> <2D bounding box in the format of [xmin, ymin, xmax, ymax]> <3D bounding box in the format of [xpos, ypos, zpos, width, height, length, angle]>

	calib_kitti
		calibration file on KITTI format. P0, P1, P2, P3 corresponds to the same intrinsic camera matrix.
	
	information
		raw data on json format. You can find all the objects and scene information in a python dictionary variable.
		
	Depth
		folder containing 640x480 8-bit images. Depth is encoded in the 3 channels using the following formula: Depth = 5000 * (R + G*256 + B*256*256) / (256*256*256 - 1)



Class		R	G	B	ID

Void		0 	0 	0	0

Sky             128 	128 	128	1

Building        128 	0 	0	2

Road            128 	64 	128	3

Sidewalk        0 	0 	192	4

Fence           64 	64 	128	5

Vegetation      128 	128 	0	6

Pole            192 	192 	128	7

Car             64 	0 	128	8

Traffic Sign    192 	128 	128	9

Pedestrian      64 	64 	0	10

Bicycle         0 	128 	192	11

Lanemarking	0	172 	0	12

Reserved	- 	- 	-	13

Reserved	- 	- 	-	14

Traffic Light	0 	128 	128	15