cpflame / loco Goto Github PK

Thank you for your sharing. The tar00 seems broken?

Your work is excellent, during the experiment.I have a few questions to ask you.
Why does your training log show solver='models/pascal_voc/VGG16/faster_rcnn_end2end/solver.prototxt' instead of contxt-solver？

Hi CPFLAME,

the link of model is gone
"For testing We released our pretrained model at model, you can download it for testing."

Could you please put pretrained model again? thanks very much.

Hi, thanks for sharing this code.

I'm interested in the BDCI dataset's detection result. Now I've got BDCI training images and annotations, and your LOCO algorithm inspires me. I'd like to reveal your reported detection result in your paper, but find that you didn't mention these settings:

• anchor aspect ratios
• anchor scales
• RPN_min
  Are they the default parameters as used in py-faster-rcnn? i.e. [0.5, 1, 2] for anchor aspect ratios, [8, 16, 32] for anchor scales? What about RPN_min for BDCI? Is it 16 pixels as in py-faster-rcnn or 5 pixels as in your TT100K experiments?

Also, I'm not quite sure, is that true that you set these parameters in BDCI detection dataset:

• iter_size = 1, instead of the default iter_size=2 (for VGG16)
• MAX_SIZE=1000, and SCALES: [1000,] for training and single scale testing

What's more, there are exactly 10000 images for training with annotated bounding boxes in BDCI16-TSDAD2, and 4000 annotated images in BDCI16-TSDAD1. How could you do experiment on BDCI16-TSDAD2 with 13500 training images and 23000 testing images, as described in your paper?

Looking forward to your reply, thanks.