flickr30k upperbound

[volkancirik]

Hello,

I used Bottom-up Attention to get boxes for Flickr30k data. Unfortunately, I could not get the same upperbound you reported in the paper. I get 0.6507 you reported 0.8745. Do you mind providing the details how you used Bottom-up model for inducing boxes. Below I listed mine:

model_name: resnet101_faster_rcnn_final.caffemodel
conf_thresh=0.2
min_boxes=10
max_boxes=100

UPDATE:

When I increase the number of boxes I get better upperbound but still it is not as good as yours, below setup gives me upperbound 0.8530

model_name: resnet101_faster_rcnn_final.caffemodel
conf_thresh=0.01
min_boxes=200
max_boxes=200

[jaesuny]

Did you train Bottom-Up Attention model yourself?
When I used Bottom-up Attention model pretrained on Visual Genome, I got upper bound about 0.87.

[volkancirik]

No I did not train. what are the hyperparameters that you use for running the model?

[jaesuny]

I just ran tools/genenerate_tsv.py of Bottom-up Attention

As default,
model_name: resnet101_faster_rcnn_final.caffemodel
conf_thresh=0.2
min_boxes=10
max_boxes=100

[upccpu]

Hi,could you tell me which kind of GPUs you used,I use two titan Xp to run this only to show 'out of memory'.Thanks a lot.

[jd730]

@upccpu They used 4 GPU(12GB each) See this