DAVIS24
DAVIS Event Camera Sample Data
DAVIS24 has unpublished samples of DAVIS event camera data, for exploring raw data and for algorithm development.
DAVIS24 samples were recorded by Tobi Delbruck, Sensors Group of the Inst. of Neuroinformatics, Univ. of Zurich and ETH Zurich.
Information about other datasets and code are on the Sensors Group webpage.
Dataset files
You may (not clear if true in 2024) be able to add a shortcut of this folder to your google drive to synchronize the files to your local drive. Otherwise you can download the files individually.
Format
Files were mostly recorded from a DAVIS346 camera with a variety of lenses, and are mostly in AEDAT-2.0 format that can be played with jAER.
Playing with the data
You can drag and drop the files onto the jAER AEViewer. Use the AEChip Davis346blue to play most of these files.
Dataset contents
See the DAVIS24 gdrive data. File names give hints to contents.
Stand by for more detailed descriptions of dataset files and sample videos
Citation
Website citation is appreciated:
T. Delbruck, "DAVIS24: DAVIS Event Camera Sample Data", Available: https://sites.google.com/view/davis24-davis-sample-data/home
Contact
Please post on https://groups.google.com/g/davis-users if you want to see more/different sample DAVIS346 data here.
Event Camera used
These samples were recorded from a prototype DAVIS346 frame-event camera, (commercially sold by inivation), specifically the one briefly reported in
G. Taverni et al., “Front and Back Illuminated Dynamic and Active Pixel Vision Sensors Comparison,” IEEE Trans. Circuits Syst. Express Briefs, vol. 65, no. 5, pp. 677–681, May 2018, doi: 10.1109/TCSII.2018.2824899. Available: http://dx.doi.org/10.1109/TCSII.2018.2824899
The original DAVIS (DAVIS240) was reported in
C. Brandli, R. Berner, M. Yang, S. Liu, and T. Delbruck, “A 240 × 180 130 dB 3 µs Latency Global Shutter Spatiotemporal Vision Sensor,” IEEE J. Solid-State Circuits, vol. 49, no. 10, pp. 2333–2341, Oct. 2014, doi: 10.1109/JSSC.2014.2342715. Available: http://dx.doi.org/10.1109/JSSC.2014.2342715
Recordings
Initial batch, 20.2.24:
'Davis346 APS+DVS samples with ring light.aedat': A deck of cards is riffled with frames and events recorded. The frames are totally overexposed and the frame rate is too low to capture the faces, but the events see most of the card vfeatures.
'Davis346blue-2018-12-07T13-56-01+0100-00000001-0 disk high BW pr.aedat': A disk with black dot rotating at about 50Hz with a pixel speed of about 18k px/s; the photorreceptor is turned up to high bandwidth to capture the leading dark edge. The trailing light edge still makes a long trail of events.
'Davis346blue-2018-12-30T04-44-31-0800-00000001-0- tobi juggling pasadena shorter.aedat': Juggling balls with a swinging rope in foreground, recorded with frames and events under low illumination (<5lux). The frames are totally blurred and there are lots of shot noise events.
'Davis346blue-2019-01-14T14-20-34+0100-00000001-0 ramp up down.aedat': The rotating disk data used for our paper comparing Nyquist-sampled and Event-sampled data (S.-C. Liu, B. Rueckauer, E. Ceolini, A. Huber, and T. Delbruck, “Event-Driven Sensing for Efficient Perception: Vision and Audition Algorithms,” IEEE Signal Process. Mag., 2019)
'Davis346blue-2019-06-18T12-28-04+0200-00000002-0 train bias control.aedat': Rain with frames and events and automatic control of refractory period to limit event rate, as reported in T. Delbruck, R. Graca, and M. Paluch, “Feedback control of event cameras,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). 2021
'Davis346blue-2019-11-28T06-51-04+0100-00000003-0 blaine card change.aedat': Frames and events from attempts to do David Blaine's instantaneous card change.
'Davis346blue-2020-02-10T10-57-27+1100-0 ping pong DVS only.aedat': DVS only from ping pong game, recorded at the RVSS in Canberra.
'Davis346blue-2020-12-02T16-45-49+0100-00000003-0 cards.aedat': Another recording of a deck of cards riffled with frames and events recorded. In this recording, the frames are underexposed so quite dark.
'Davis346mini-2017-02-27T12-40-18+0100-00000002-0 DVS only flow gravel IMU 250dps.aedat': Frames and events of gravel; maybe useful for dense optical flow. The IMU is also recorded and can be used to perhaps help fuse IMU and OF.
'Davis346mini-2017-04-29T08-33-14+0200-00000004-0 ben and konstantin CNE.aedat': Father and child, DVS only.
'Davis346mini-2017-04-29T16-04-39+0200-00000004-0 CNE bar 2017.aedat': Bar at the capocaccia workshop; frames are datadriven by accumulation of a particular number of events or event rate exceeding a threshold, via the jAER filter DavisAutoShooter.
'Davis346mini-2017-05-15T16-27-56+0200-03460001-0 3p5 mm gyro calib.aedat': Stationary camera data to calibrate the camera IMU offsets.
'Davis346mini-2017-05-24T15-53-18+0200-0 office run relogged.aedat': A DVS only recording of a run through INI.
'Davis346mini-2017-06-29T17-26-58-0600-03460001-0 ball sand 2.aedat': Frames and events recording of a beach volleyball kicked between people, recorded in Telluride.
'Davis346mini-2017-07-10T14-36-38-0600-03460001-0 sb2.aedat': DVS only recording of a superball bouncing on a table and then falling off; recorded in Telluride.
Davis346mini-2017-11-02T12-05-19+0100-00000001_card_throwing.aedat: Throwing spinning cards into the air, high noise level under low lighting.
'Davis346mini-2017-11-02T15-33-03+0100-00000001-0 nccr girls.aedat': Two young ladies clowning around for the camera; DVS only.
'Davis346mini-2017-11-16T10-15-54+0100-00000001-0 fast disk.aedat': Another fask rotating disk with a dot on it, frames and events.
'Davis346mini-2017-11-20T10-46-33+0100-00000001-0 hand wave dark.aedat': Waving a hand around under very low illumination, with high shot noise rate.
'Davis346mini-2017-12-19T16-04-50+0100-00000001-0 roshambo samples tobi.aedat': Rock scissors and paper hand symbols, useful for testing jAER RoshamboCNN and python Dextra.
'Davis346mini-2018-09-24T15-53-56+0200-00000001-0 2 coins.aedat': Two coins that are spun on a plate and collide with each other.
Note: for above recordings, the DAVIS346 IMU settings were mostly default values of full scale 250 deg/s gyro and full scale 2g accerleration; set jAER to these in the Hardware Configuration panel to get correct IMU values.
Added 23.2.24: Telluride Mountain Biking Data
See the mountain biking gdrive folder.
2018 Telluride Neuromorphic Workshop recordings with Alex Zhu from mountain biking with DAVIS336 mounted on handlebars of bike; laptop was in backpack. 25 FPS frame rate for gray frames. DVS/APS/IMU samples. Exported snippets from source rosbag to AEDAT-2.0 files.
Ask on davis-users forum if you want the complete 1500s recording in rosbag format (62GB).
These files are big and so supplied as highest compression level 7-zip files with 7z a -t7z -mx=9, which achieves about 7X compression.
village_trail_2018-07-12-13-03-27-start-IMU-calib.aedat.7z: Stationary bike, useful for checking IMU accelerometer calibration, which should show about -1g downwards.
village_trail_2018-07-12-13-03-27-start-of-trail.aedat.7z: Starting out from gondola.
village_trail_2018-07-12-13-03-27-entering-forest.aedat.7z: Entering the forest from sunny trail.
village_trail_2018-07-12-13-03-27-forest-single-track.aedat.7z: Deep in the forest single track trail.
village_trail_2018-07-12-13-03-27-paved-path-car-crossing.aedat.7z: Riding on paved pedestrian path and crossing at crosswalk; car passes in front.
village_trail_2018-07-12-13-03-27-mountain-village-plaza-pedestrians.aedat.7z: Coming into Mountain Village plaza; ends with some distant pedestrians.
village_trail_2018-07-12-13-53-59-riding-behind-alex-single-track.aedat.7z: Riding behind Alex on single track trail in forest.
Note: When playing these mountaing biking AEDAT, set the correct IMU full scale values in jAER as shown to right, or the IMU values will not make sense. When these AEDAT files were exported, these settings were used and they must be identical full scale gyro and acceleration values (2000 deg/s and 8g) to see correct values when playing back the data.
Since the IMU was mounted on handlebars, there are places where acceleration exceeds 100g according to ROS bag values (it doesn't really make sense because the full scale that the DAVIS346 can measure is 16g according to the datasheet of the Invensense MPU-6150a IMU.
Want more samples?
Post on https://groups.google.com/g/davis-users to ask.
Want other kinds of data?
Easy-to-access sites
The Event-Camera Dataset - rosbags and CSV files, pure DVS, DAVIS240C camera, simple scenes
MVSEC - rosbag Multicamera dataset, can play DAVIS from left channel by dropping onto jAER
DND21 - AEDAT pure noise recordings and synthetic or cleaned-up real scenes from driving, surveillance, and moving particles. Go directly to AEDAT folder.
DVS09 - AEDAT recordings of DVS128 (pure DVS, the first commercial event camera)
All above datasets are verified to play in jAER; just drop the .bag or .aedat file onto the AEViewer window or use File/Open...