search

shizenglin / training-free-object-counter

44 stars 3 forks source link

do you compare with grounding-sam , it seems that prob not better than grounding-sam #3

Open senlin-ali opened 1 year ago

shizenglin commented 1 year ago

what is grounding-sam and what do you mean by saying "prob not better than grounding-sam"? sorry, cannot fully understand your question. Could you expalin more?

senlin-ali commented 1 year ago

eeee... you can see this https://github.com/IDEA-Research/Grounded-Segment-Anything

shizenglin commented 1 year ago

Prior to proposing our new method, we experimented with grounded DINO on the FSC147 dataset and found that it frequently detects all objects of the same class using a single large bounding box, especially when dealing with small objects. It would be great if you can revisit this approach and kindly share your results here.

rruiz-s commented 9 months ago

In an experiment using Grounding DINO and Grounded-SAM with a dataset containing (69 x 47 px) objects, the results are consistent with @shizenglin 's findings. For clarification, instead of identifying a single animal (the "object"), both (Grounding-DINO and Grounded-SAM) identified the whole herd as a single large bounding box. Based on this, there is a need of new methods that tackle these challenging scenarios.