search

linjieyangsc / densecap

Dense captioning with joint inference and visual context
Other
53 stars 24 forks source link

Predict captions for existing bounding boxes #5

Open idobrusin opened 5 years ago

idobrusin commented 5 years ago

Hi,

we would like to use densecap to predict caption for already computed bounding boxes. I tried using the im_detect() function in /lib/fast_rcnn/test.py which has a boxes argument.

I would expect the function to output the same number of boxes as i put in, which does not happen. Instead, it looks like the network predicts new boxes in the RPN.

I tried setting the cfg.TEST.HAS_RPN parameter to False in order to load the rois blobs in the _get_blobs() function -> The boxes are loaded into blobs, but this has no effect on the outcome. Are they used at all in this case?

Do i need to adjust the feature_net (vgg_region_global_feature.prototxt) in some way or set some other parameters in order for the network to work as expected? Or did I miss something else?

Thanks

alice-cool commented 3 years ago

Can modify proposal_layer.py 's forward function image