about baseline in paper
dangnh0611 opened this issue · 1 comments
dangnh0611 commented
Hi, thanks for your works. When reading your paper, i have few questions about the baseline:
- Baseline is build based on CenterNet with an extra head predicting tracking offset map
O_B
. As mention in #36 , backbone network just take a single image (frame at timestamp t) as input instead ofprevious_frame
+current_frame
+previous_detection_heatmap
like CenterTrack? So I wonder how to predict offset mapO_B
with just one frame. My initial guess is that feature map of framet
andt-T
is calculated independently, but attaching offset-map head on an aggregated feature map (e.g simple stack/average/subtraction) oft-T
andt
? - In Fig. 4, since Baseline doesn't have
O_C
(w/o CVA), does it meansBaseline == Baseline + CVA
orO_C == O_B
in the baseline context?
JialianW commented
Thanks for your interest! The baseline in the paper should refer to just adding a simple tracking head on CenterNet. The O_B in baseline should be predicted from the single current image only.