I have a question about heatmap

[jaekyu-sim]

I'm very thanksful about your github page.

When I see your github page about Pose_Estimation project,

I saw that we need heatmap 19(because 18 joint + 1 background)

but I can understand why I have to consider background

[hellojialee]

Also interested. What the performance will be if we drop the background channel?

[michalfaber]

@Jaekyu-Sim @USTClj I’ve trained the model without background channel and it looks like it doesn’t affect the detections. The loss values are slightly smaller. I thought that the background could be helpful in calculating coordinates - non maximum suppression around the lines: all_peaks = [] peak_counter = 0 for part in range(19-1): …but authors of original solution don’t use the background channel - maybe it was a result of some previous experiments. I noticed that when I add the background I get twice as many peaks - most of them don’t affect the final result and can be discarded.