search

piotrhm / image_recognition_project

0 stars 0 forks source link

Prototypes colors experiment #30

Closed mproszewska closed 2 years ago

mproszewska commented 2 years ago

Results are not great, but at least show that our visualizations make some sense. Visualizations depend on colors of prototypes.

Besides, IMO prototypes that are found during training are not as interpretable as one might want. I base that hypothesis on bounding boxes visualization, which for Red bellied Woodpecker make some sense (2 distinguishable prototypes), but not so much for other birds. Note that our bounding boxes found for Florida Jay differ from the one in the paper. Maybe that's a bug.

DorotaNowak commented 2 years ago

From what we have those are the best visualizations that actually make sense. I also wouldn't expect much from those prototypes - it looks like for Red Bellied Woodpecker there are two that the model distinguishes (white-red head and black feathers) and for Yellow Breasted Chat apparently there is only one (yellow belly). And they might've used different architecture than our resnet and that's why the bounding box for Florida Jey is different.

mproszewska commented 2 years ago

Can we merge it? @piotrhm

DorotaNowak commented 2 years ago

Say yes @piotrhm.