I am doing some testing using the trap.nz website to upload images into a Monitoring project. We would love to be able to use your system to help us protect kiwi in our area.
I have done a whole bunch of testing with the same 128 image test dataset. In my testing I have found:
The duplicate detection does not work. It detects duplicates that are not duplicates (2 of 128 in my test dataset were detected as duplicate), and doesn’t detect duplicates that are a re-upload of the exact same image file (when I re-upload the same 128 images, 76 images were found to be unique, 52 duplicate).
The ‘Date’ listed seems to round the time into the nearest 15-minute window. So all images are listed with a datetime of: dd mm yyyy hh:00, dd mm yyyy hh:15, dd mm yyyy hh:30 or dd mm yyyy hh:45.
Is this by design? I see the original EXIF data on the image is stored correctly after upload. It makes it hard to check if you have uploaded an image already as we can’t see the original filename nor (easily see) the correct timestamp, so the trap.nz ‘Date’ field is all we have go on.
- If there is no predator detected, there is no way to tell that the AI has actually run. If an image is scanned by the AI and nothing detected, it should flag ‘Specifies detected’ as ‘None’ and set a confidence level so that you know the AI has analysed the image and found nothing.
If the AI algorithm doesn’t give a confidence level when there is no predator detected, just set it to 0%. That way we can just tick the reviewed box rather than having to (a) wonder if the AI has run and (b) select None on a whole bunch of images manually.
(I am happy to provide details of my testing/dataset if required for troubleshooting by developers).