ISETHDR CV Experiment

Methods

Image generation

First, we need to assemble a dataset of driving images to run the YOLO algorithm on. We acquired four driving scenes from the ISR HDR Sensor Repository. Each scene includes .exr files that contain radiance data for the sky, street lights, headlights, and other lights.

Scene images <insert images> 1112184733 1112201236 1114031438 1113094429

When setting up the scenes, we consider four lighting scenarios that commonly occur while driving. The light scenarios are described by a weight for each .exr file.

<insert table> Day - strong sky illumination [0, 0, 0, 50]

Dusk - weak sky illumination with headlights and streetlights [0.2, 0.001, 0, 20]

Night - headlights and streetlights [0.2, 0.001, 0, 0.0005]

Blind - extreme headlights and streetlights [2, 0.1, 0, 0.0005]

Next, we set up the camera. We select four sensors:

That sets up the scenes. Now we set up the cameras.

We use 4 different sensor types: three automotive and one smartphone.

Automotive sensors

ar0123at: 1.2 MP, dynamic range >115 dB mt9v024: 0.36 MP, dynamic range >100 dB ov2312 split pixel: 2 MP, dynamic range >68 dB

Smartphone sensor imx363: 12 MP

Now build the data set

use exposure time to alter dynamic range

times = 0.1, 0.5, 1, 2, 4, 6, 8, 12, 16, 20, 50, 100, 500, 1000 ms

Loop over: 4x scenes 4x lighting scenarios 4x sensors 14x exposures

total images - 896

YOLO

Appendix