Signals Dataset

This dataset is used to traing a YOLOv3 object detector on signals with different colors and numbers on them.

used for https://github.com/eddex/pren2

annotations

bounding-boxes-digits

Bounding boxes around digits, no distinction between different colors.

Total number of annotations: 540

class	1	2	3	4	5	6	7	8	9	start
annotations	32	18	63	81	62	68	113	60	31	12
distribution	5.9%	3.3%	11.7%	15.0%	11.5%	12.6%	20.9%	11.1%	5.7%	2.2%

bounding-boxes-digits-infolks

Bounding boxes around digits, no distinction between different colors. Annotation done by INFOLKS.

Total number of annotations: 5687

class	1	2	3	4	5	6	7	8	9	start
annotations	534	494	716	612	690	503	779	617	456	286
distribution	9.4%	8.7%	12.6%	10.8%	12.1%	8.8%	13.7%	10.8%	8.0%	5.0%

the data

There are 37 different signals.

• 9 Number signals with white font and black background on a high pole (info-signals)
• 9 Number signals with black font and white background on a high pole (info-signals)
• 9 Number signals with white font and black background on a low pole (stop-signals)
• 9 Number signals with black font and white background on a low pole (stop-signals)
• 1 start signal: blue and white striped signal on high pole

camera settings

camera = PiCamera()
camera.exposure_mode = 'off'
camera.resolution = (416, 416)
camera.brightness = 55
camera.framerate = 15
camera.shutter_speed = 3100
camera.rotation = 0
camera.iso = 1200

how the data has been collected

• The train was set to full speed (7) and slow driving (2) while taking images.
• The signals have been spilt to 2 sets because not all could fit on the track at the same time.
• Batches of 1000 images have been taken while driving.
• The signals positions were (sometimes) shuffled between batches.
• Images without any signals have been deleted afterwards.

Does this look like a scientific approach on collecting data? WARNING: The answer might be shocking!

first set of signals

signals

• info white: 2, 4, 5, 6, 8
• info black: 6, 7, 8
• stop white: 1, 3, 4, 5, 6, 7, 9
• stop black: 1, 3, 4, 6, 7, 8, 9
• start

collected data

• batch 0 - 999: speed 7
• batch 1000 - 1999: speed 2
• signals shuffled
• batch 2000 - 2999: speed 7
• batch 3000 - 3999: speed 2

second set of signals

signals

• info white: 1, 3, 7, 9
• info black: 1, 2, 3, 4, 5, 9
• stop white: 2, 8
• stop black: 5
• start

collected data

• batch 4000 - 4999: speed 7
• signals shuffled
• batch 5000 - 5999: speed 2
• signals shuffled
• batch 6000 - 6999: speed 7
• signals shuffled
• batch 7000 - 7999: speed 2

third set of signal(s)

signal(s)

• stop black: 2

Because I missed this one...

The signal has been placed in different positions between and during the 4 rounds.

• batch 8000 - 8999: speed 7
• batch 9000 - 9999: speed 2
• batch 10000 - 10999: speed 7
• batch 11000 - 11999: speed 2