About downstream task semantic segmentation on DSEC

[Candy-Crusher]

May I ask how did you process the event input of the ResNet-34 backbone and Deeplabv3 decoder? What is the bin size that you set?

[jeongyh98]

Our model translates a real-day event histogram into a night event histogram. The shape of the night event histogram is similar to that of RGB images or voxels, which are typically structured as C x H x W. For example, in a 3-bin event histogram, there are 6 channels, with each bin having a positive polarity channel and a negative polarity channel. To accommodate this, we modify the first layer of the backbone to accept 6 channels for the event encoder. For 1-bin event histogram, we set as 2 for input encoder. The decoder retains the same layers as in the original model because the form of the result must be the same.

We are in the process of updating how we pre-process raw events into event histograms.