question about output

[lzhengchun]

How can I interpret your output? I noticed that the output has three channels.

I assume this is not object localization.

[liponan]

Hello! They are confidence score, y_offset, x_offset.
Let's say at pixel [20, 40], these three channels give you 0.75, 0.12, 0.26 (after Sigmoid activation).
That means that it predicts there is a peak with 0.75 confidence at (40.26, 20.12).
Note that by "offset" it means the relative offset within a pixel.

[lzhengchun]

Thanks for your response! what kind of processing are further required on the model output to get the peak location with sub-pixel precision ? for example, a peak is composed with multiple connected bright pixels, i assume the model with give high confidence. do you use connected component labeling and use the maxima point as peak?

[liponan]

Sorry for my delayed response. I use non-maximum suppression (NMS) to remove false positives, i.e. those pixels connected to the true central pixel. The idea is we sort detected peaks by their confidence score, then add them to a new list one by one. If the peak to be added is within a certain cutoff distance to any other peaks already in the new list, then we drop it.

[lzhengchun]

Sorry for my delayed response. I use non-maximum suppression (NMS) to remove false positives, i.e. those pixels connected to the true central pixel. The idea is we sort detected peaks by their confidence score, then add them to a new list one by one. If the peak to be added is within a certain cutoff distance to any other peaks already in the new list, then we drop it.

oh I see. thanks