search

facebookresearch / maskrcnn-benchmark

Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.
MIT License
9.29k stars 2.5k forks source link

Reducing the size of predicted bounding boxes. #944

Open mtlouie-unm opened 5 years ago

mtlouie-unm commented 5 years ago

❓ Questions and Help

Is there a way to reduce the size of the predicted bounding boxes? For example, I am using the 2018 xView satellite imagery dataset. After I train a model and perform inference on a test set, the predicted bounding boxes are much larger than the actual target. I have tried reducing the _C.TEST.BBOX_AUG.MAX_SIZE parameter in the defaults.py from 4000 to 500, but I don't see any changes when doing the inference. Are there other parameters I need to change in defaults.py?

mtlouie-unm commented 5 years ago

Here is an example of what I am running into. The blue boxes are the annotations and green boxes are the predicted bounding boxes. As you can see they are quite large.

253102_pred
rnsandeep commented 5 years ago

Change the rpn anchor sizes. Try to decrease the size as less as 4 and see.

hunaodaotianming commented 4 years ago

I had the same problem. Did you solve this problem?