Any point to reducing pos_dist_thresh over time?

Box 2 of “” describes pos_dist_thresh as

all locations within this distance threshold in pixels are considered positive training samples for the detector… For very small objects, such as a fly leg in a large image, you should consider lowering the value.

Intuitively, it seems like it might be worth starting with a large distance threshold (so that the network hits at least some positive examples) and then gradually making the threshold stricter over time. I am curious as to whether non-static values have been tried previously.

1 Like

I think that is a good idea, and could improve performance a bit. One issue with starting with too small distance thresholds is that the convergence is slow (see Supp Mat Fig 2). So your adaptive approach could help. However, due to the locref layers (which counteract the discretization errors due to downsampling along the ResNet), the effect might not be too big. I.e. the performance is pretty robust over broad ranges of pos_dist_thresh, as this figure from Supplementary Figure 2 in shows.