persts / bboxee Goto Github PK

In issue #3 the idea of adding the ability to flag an image for review. I can see the value in this idea, but would like to see if others have input on just how this would be used.

feature that allows tracking of individuals with identifiable markings (ex, bobcats w/ spots)
--would this just be creating new classes for individual animals or an entire separate feature?

there is a feature in CPW PhotoWarehouse that describes the species in terms of an overarching "Group" (ex, ungulate for elk) and you can incorporate further "Details" into a particular species identification (ex, calf for elk). Is there a way to incorporate this information when transferring this kind of data over if we use, for example, hierarchical modeling?

Ex:
(Species Model) --> "Elk" label
(Elk Details Model) --> "Calf" label
(Transfer to PW) --> Ungulate Group, Elk Species, Calf Detail
**group would be filled out automatically since each species can only be one group

feature that reports model's confidence level for bboxes created with the assisted annotation tool
-incorporate an additional feature to "flag" boxes below a preset threshold of confidence for manual review?

Using BBoxEE.exe beta 3, when attempting to export a a couple batches it crashes. See attached screenshots for .cmd notes.

Notes:
4000 total photos
<2000 bounding boxes
4 class types
no exclusions for truncated/etc.

Hi! I like the app, I find the auto-annotator very neat. I am reviewing existing tools for annotation camera trap data, and I was wondering whether you had plans to add features to the app, and in general what is the status of development on that project.

something that allows BBoxEE to assume a blurry photo's class assignment based on the other photos within the same trigger event.

Ex: [elk passes through camera site]
IMG01 = elk on left
IMG02 = elk at center
IMG03 = blurry (elk right up next to camera lens)
ImG04 = elk on right
--> BBoxEE assigns "elk" to IMG03 based on assignments for IMG01, IMG02, and IMG04

Considerations:

1. Would this require use of the photo metadata to read timestamps and identify the photos that are within the same event, or would this be done by assuming the photos immediately adjacent to the "unknown" photo are relevant?
2. How would this resolve if the photos used are not equal? Ex, [2 elk] [unknown] [1 elk]?
3. I assume since the AI cannot distinguish a "blurry" definition very intuitively, measuring low confidence would be a better approach?

Some sort of method to track accuracy of the imported model once the assisted annotation has been used on a dataset

Possible option:
-track changes to bboxes (adjustments to bbox dimensions and changes in class)
-provides a % and real number of bboxes adjusted, and divided by dimension changes versus class changes
-possible to incorporate a type 1 and type 2 error rate for whole batch? Per class in batch?

When I go to export my data, sometimes I find a technician has forgotten to label the boxes they drew for a batch, and it's left me with multiple NA labels across multiple .bbx files. I'm not able to simply relabel the NA to the correct species in the export tab because it will often include multiple species. Would it be possible to create a "Skip to Next NA Labelled Photo" button to the Annotate Tab, similar to the "Skip to Next Photo Containing a Bounding Box"? Then I wouldn't waste an hour searching each .bbx that has an NA label somewhere in it.

Hiya,

I don't have any issues, well, I have loaded the interface but have no idea how to load images and how to use the application for camera trap images.

Any further info you can provide I will be eternally grateful

All the best
Mark

bboxee will sometimes crash during manual annotation. Not yet found a common denominator, but it occurs on my work computer (desktop hooked into a high-security network server) and on an employees laptop (standard laptop not connected to any network).

Only pattern is crashing after at least a few hundred photos have been labeled. Does not occur every time, or even most times, or on particular days, or while anything else is/isn't running simultaneously. Very unpredictable when it will crash. Does not corrupt the .bbx file, but loses any unsaved progress.