search

agentmorris / MegaDetector

MegaDetector is an AI model that helps conservation folks spend less time doing boring things with camera trap images.
MIT License
116 stars 24 forks source link

class labels significance #16

Closed agentmorris closed 1 year ago

agentmorris commented 1 year ago

Where can I get the mapping between class labels and class names in "run_tf_detector.py" ?

Issue cloned from Microsoft/CameraTraps, original issue posted by saikatdutta on May 19, 2019.

agentmorris commented 1 year ago

Hi @saikatdutta, please see experiments/megadetector_v3/label_map.pbtxt for the class label mapping for the MegaDetector (version 3, to be released shortly; in version 2 there is only one class, 1:animal).

(Comment originally posted by yangsiyu007)

agentmorris commented 1 year ago

OK, but when I ran the code on megadetector_v2, I got some labels other than 1(55,83 etc.). What are those labels actually?

(Comment originally posted by saikatdutta)

agentmorris commented 1 year ago

The megadetector was warm started with COCO weights and we left all the layers in, so the output technically has 90 classes (the COCO stuff classes). But the model is only meant to be applied to camera trap photos, and only class 1 (and when we release v2, class 2 as well) should be considered meaningful.

(Comment originally posted by yangsiyu007)