Wild Dog Detector mislabelling tails as bodies

In which Wildbook did the issue occur? ACW

What operating system were you using? (eg. MacOS 10.15.3) Win 10

What web browser were you using? (eg. Chrome 79) Chrome latest

What is your role on the site? (admin, researcher, etc) Admin

What happened?
Latest bulk upload has been run and on beginning matching process the system is presenting tails for matching - on inspection of the annotations - many (not all) are not labelled as tails and hence the system is attempting to match.

What did you expect to happen?
a. Tails be labelled as tails
b. Matching then working properly on bodies

What are some steps we could take to reproduce the issue?

Se the following Sightings and associated encounters in ACW.
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=f76a810a-6634-404e-87f6-61c4db7d6bb5
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=22a891dc-6310-4605-92de-57c83c18980a
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=daf50190-0de1-4649-b29b-b36d0dfd3da6
https://africancarnivore.wildbook.org/encounters/encounter.jsp?number=4bfda2e9-1b63-466b-a32b-470ebb9be468

These are just a sample of the issue

Not all Sightings/encounters seem to generate the problem.

If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report

@jason
@MarkF
@ACWadmin1

Hello,

I’ve been looking at this today. It looks like these are incorrect classifications from the model itself. There have been no changes to the model and config, and from my observation sample there are mostly correct class decisions. The bodies of dogs seem to be having matching applied correctly.

I can certainly verify the incorrect classes on the above four encounter links, but was unable to find any others in the import looking through twenty or so encounters and scrolling through the import in the job log. Are there many more?

Since it is a true classifier error, the only way to prevent it in the future is to retrain with more diverse tail data. I’d suggest deleting the misclassified tail encounters when you find them and manually re-adding the tail annotations if it’s helpful to do so.

Thanks!

Thanks Colin

I think there is more to this than meets the eye for the following reasons.

  1. This has not been an issue prior to the last few weeks - prior uploads, although not perfect on tail detection, never mislabelled a tail as a body.
  2. It seems to occur in clusters within Sightings - some some Sightings have no such issue and others do - and I cannot see a pattern to this.

As I work thru this Census data (a further 530 Sightings to go) I will record the instances I discover and once I have a large number of examples - will send to you and JP for analysis.

Thanks

Paul
@ACWadmin1
@colin
@parham

1 Like