Visualizing bbox predictions/proposals on images (Tensorboard) during training?

[cpereir1]

Hi!

Has anyone implemented visualizing the predictions (bbox in this case) overlaid on images, for display for example in Tensorboard, while training/fine-tuning a detectron2 model using train_net.py (for example)? This would be very useful to have, and see it is not there out-of-the-box. Hoping someone has had the same need and could share his/hers implementation.

Many thanks and great work so far!

[sukjunhwang]

I believe the function below can help you.

detectron2/detectron2/modeling/meta_arch/rcnn.py

Line 87 in 31ec19b

def visualize_training(self, batched_inputs, proposals):

[Shuntw6096]

You can set VIS_PERIOD in config file, it will do visualization every few iterations during training.

detectron2/detectron2/config/defaults.py

Line 624 in 0e6702d

_C.VIS_PERIOD = 0