Open allenleetc opened 6 years ago
Related to this, any thoughts how can we show confidence/scores for the predictions?
On Wed, Mar 7, 2018 at 4:10 AM Allen Lee notifications@github.com wrote:
Idea is an automated confidence/suspiciousness measure for CPR tracking results. For data where the base tracking is pretty good, the replicate dispersion tends to be higher for more inaccurate frames as intuitively makes sense. In addition to this one can add a trajectory smoothness constraint. Combining these two factors might net a decent warning indicator about the tracking looking "pretty wrong" or "pretty right" at certain frames.
Besides as a general check for tracking performance, two possible usages for training data augmentation: i) APT suggests "this frame looks suspicious" to user for (re)labeling. Intuitively it seems that targeted labeling/augmentation in this way would be incrementally better than augmentation by labeling frames randomly. ii) APT suggests "this frame is probably right" to user for easy/rapid acceptance. Here the idea is that "verification is easier than labeling from scratch" which seems plausible esp when there are a larger number of landmarks.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kristinbranson/APT/issues/185, or mute the thread https://github.com/notifications/unsubscribe-auth/ABLVY0R4Qn-4YLj58YMfPHEynBxDe84uks5tbxBNgaJpZM4Sfjz1 .
Hmm yea I guess it's either on the markers themselves or in the axes below? Does your confidence come as a spatial uncertainty (eg units of px) or would it be an arbitrary number? I think for CPR it's prob just a number with arbitrary units for now. Maybe it can always be mapped to a spatial uncertainty, if only empirically etc via a 'calibration'. Just brainstorming that could be a stretch.
If it can be mapped to spatial units we could use some kind of spatial indicator that is faithful to the actual scale (on/off toggleable). Another idea is to show something during a mouse-hover-over-marker? Just making regular plots in the axes below seems good too since it gives the temporal context. What do you think?
I'm not going to use it, so might be better to ask Stephen and Alice about it. I'll move this discussion to slack.
On Thu, Mar 8, 2018 at 2:52 AM Allen Lee notifications@github.com wrote:
Hmm yea I guess it's either on the markers themselves or in the axes below? Does your confidence come as a spatial uncertainty (eg units of px) or would it be an arbitrary number? I think for CPR it's prob just a number with arbitrary units for now. Maybe it can always be mapped to a spatial uncertainty, if only empirically etc via a 'calibration'. Just brainstorming that could be a stretch.
If it can be mapped to spatial units we could use some kind of spatial indicator that is faithful to the actual scale (on/off toggleable). Or we could show it during a mouse-hover-over-marker? Just making regular plots in the axes below seems good too since it gives the temporal context. What do you think?
— You are receiving this because you commented.
Reply to this email directly, view it on GitHub https://github.com/kristinbranson/APT/issues/185#issuecomment-371289160, or mute the thread https://github.com/notifications/unsubscribe-auth/ABLVY9vJ_L054Ub8LhyCQjHSRfSkrg6Yks5tcE-QgaJpZM4Sfjz1 .
Late to party, but mayank mentioned thinking about things to add 'navigation" options for APT and axes display, and I think 'low' confidence frames would be a good one. I think options to showing/navigate to the min confidence of all the landmarks or of a specific landmark would be helpful.
Idea is an automated confidence/suspiciousness measure for CPR tracking results. For data where the base tracking is pretty good, the replicate dispersion tends to be higher for more inaccurate frames as intuitively makes sense. In addition to this one can add a trajectory smoothness constraint. Combining these two factors might net a decent warning indicator about the tracking looking "pretty wrong" or "pretty right" at any given time.
Besides as a general check for tracking performance, two possible usages for training data augmentation: i) APT suggests "this frame looks suspicious" to user for (re)labeling. Intuitively it seems that targeted labeling/augmentation in this way would be incrementally better than augmentation by labeling frames randomly. Note the "Suspicious Frames" fcnality already exists but that is currently based on traj smoothness only. ii) APT suggests "this frame is probably right" to user for easy/rapid acceptance. Here the idea is that "verification is easier than labeling from scratch" which seems plausible esp when the number of landmarks is larger.