Datasets

CVPR2016 Sequences

We provide here the datasets which have been used in our work [1] for comparison and evaluation. The scenes are taken by a moving DVS camera (rotation+translation) and contain moving objects to highlight the capabilities of our method. All sequences have been taken with the same bias setting for the event camera (see below). The events are provided in a binary-format, which is described below.

30 Face sequence

Recommended start time: 3.15s

events_0061 Ball sequence

Recommended start time: 4.15s

events_0090 Jumping sequence

Recommended start time: 4.8s

events_0731 HDR sequence

Recommended start time: 4.15s

NOTE: Scene contains a monitor with low refresh rate, which causes many events to fire.

Format

The events are sequentially stored in a binary file. Each event is encoded into 8 bytes, where the first 4 bytes stores the absolute time-stamp in microseconds and the last 4 bytes stores the event-polarity (1 for positive, 0 for negative) and the x and y position in the image. The latter bytes can be processed by using the the following bit-masks:

image_position_x = data & 0x000001FF

image_position_y = (data & 0x0001FE00) >> 9

polarity = (data & 0x00020000) >> 17,

where data is an integer, which contains the last 4 bytes of an event.

DVS128 Biases

A detailed description of the biases can be found on the iniLabs website. We use the same bias settings as used in the jAER project.

CAS: 54
INJ_GND: 1108364
REQ_PD: 16777215
PU_X: 8159221
DIFF_OFF: 132
REQ: 159147
REFR: 6
PU_Y: 16777215
DIFF_ON: 482443
DIFF: 30153
FOLL: 52
PR: 4

[1] P.A. Bardow, A.J. Davison, S. Leutenegger, Simultaneous Optical Flow and Intensity Estimation from an Event Camera, Computer Vision and Pattern Recognition (CVPR), 2016