imageomics

/

bioclip

like 42

Follow

HDR Imageomics Institute 66

I have been working on pybioclip’s custom classifier on zebras. For testing purposes, I started with a very small dataset of just 10 images of zebras [containing 25 zebras : 18 plain zebras and 7 grevy’s].

1. First, I ran the YOLOv8 to predict the bounding boxes [which it does correctly].
2. Then on each bounding box, I applied the pybioclip’s custom classifier with 2 different set of labels [ In bioclip, the common name for plain zebra is half-maned zebra ]
3. So I tried with labels : [“Grevy’s Zebra”, “Plain Zebra”] : With this configuration, the pybioclip classifies each zebra correctly [18 plains, 7 grevy’s]
4. Then with labels : [“Grevy’s Zebra”, “Half-Maned Zebra”] : With this configuration, the pybioclip classifies the plains zebra correctly, but nearly half the grevy’s zebra are misclassified as plains/half-maned. [21 plains, 4 grevy’s]

The data and results can be found here: https://drive.google.com/drive/folders/16DJx6wHZIg07PpomJEyVyNdupTZdaZVV?usp=sharing

I am interested to know why and how this is occurring. Thank you!