on_foot is a model prediction, not a verified label. It’s correct roughly 90% of the time, so expect occasional misclassifications. Treat it as a strong signal rather than ground truth.
Feedback welcome — let us know how it works for your use cases!
Great work on this @caglarpmeta ! I’m also excited about the accuracy improvements that have recently landed to push this into the ~90% accurate territory.
Note that this feature is also available in app filters (and will match results returned by the API)
thank you for sharing, the classifier is clearly mistaken on these cases. unfortunately the bicycle/pedestrian distinction is the hardest to make and as a result we have lots of recall losses on bicycle cases like these.
While I usually appreciate other people’s effort, I am also baffled why are you so stubborn to use AI to classify the mode of transportation a picture was taken with. You know, there are much easier and simpler methods to make this classification. For me, it is a total waste of effort compared to the effect you are trying to achieve. AI is for problems that are difficult to solve with other traditional/known methods. Every tool has its purpose. What you are doing is like using a sledgehammer to fix a wrist watch or a pair of tweezers to swap a wheel on a car. Just give us a speed filter and call it a day.
Hi @GITNE - for clarity, this updated classifier is using both AI/CV and heuristics like speed to obtain a stronger signal. It is not purely a speed filter (that was an earlier version which we iterated on because we were also not happy with the quality - props to @caglarpmeta for the quality work)
We actually tried everything possible not to use AI for this, due to the large scale of the problem
Unfortunately there are no reliable enough signals that is confidently correct in all cases (for instance GPS can be very noisy in dense urban environments). In fact without any machine learned solution the accuracy peaked under 75% which is not high enough to be a reliable signal. We had to employ CV algorithms to push it above 90%.
Have you considered to let or enable contributors to tag uploads or sequences?
While it may be an impressive figure in the CV space, you do realize that this is still very far from being good enough. You need like 99.9999% or one error in a million certainty for this function to have any meaningful value to data consumers. Contributor tagging is not perfect either but I am pretty sure that the certainty rate would easily exceed 99.9%.
If you really want to spend some AI developer energy meaningfully on an also very hard problem but useful solution then try to remove outliers and noise from point clouds. I would really love to see this problem solved!
I would really love to see this problem solved!