About original CenterNet performance on DOTA about bbavectors-oriented-object-detection HOT 11 CLOSED

I didn't remove the negative images from the output of image split algorithms. For center+wh+theta, I used the same training skills and architectures except for the output branches for a fair comparison.

from bbavectors-oriented-object-detection.

I also wonder how to adopt CenterNet by adding an extra angle prediction branch, could you mention the code of this part?
Thanks @ yijingru @Tarazed

from bbavectors-oriented-object-detection.

Change heads in the main.py:

    heads = {'hm': num_classes[args.dataset],
             'wh': 2,
             'reg': 2,
             'angle': 1
             }

and change the ground-truth accordingly (wh is the width and height of the box, angle \in [-90,0) is theta in Fig1a, note ground-truth theta is already calculated in datasets/base.py. I used SmoothL1 loss for angle regression.

from bbavectors-oriented-object-detection.

Change heads in the main.py:

    heads = {'hm': num_classes[args.dataset],
             'wh': 2,
             'reg': 2,
             'angle': 1
             }

and change the ground-truth accordingly (wh is the width and height of the box, angle \in [-90,0) is theta in Fig1a, note ground-truth theta is already calculated in datasets/base.py. I used SmoothL1 loss for angle regression.

I have found a strange phenomenon: when I used the 3rd channel of 'wh' head to predict angle, I got 0.636 of mAP, but when I used an extra branch to predict angle as you said, I got 0.530 mAP... I set the random seed to 0, the performance of which is a little bit lower than random seed. I'm confused as extra branch should be better.

from bbavectors-oriented-object-detection.

Maybe training loss matters. I used weighted loss asloss = hm_loss + wh_loss + off_loss + 0.1*angle_loss. The multi-tasks are hard to optimize sometimes. But we do see the angle is hard to learn for the network.

from bbavectors-oriented-object-detection.

Maybe training loss matters. I used weighted loss asloss = hm_loss + wh_loss + off_loss + 0.1*angle_loss. The multi-tasks are hard to optimize sometimes. But we do see the angle is hard to learn for the network.

That's awesome! The loss weight of multi-task can significantly affect the results! After changing the weight of angle_loss, I got 0.68mAP! Thank you very much!

from bbavectors-oriented-object-detection.

Maybe training loss matters. I used weighted loss asloss = hm_loss + wh_loss + off_loss + 0.1*angle_loss. The multi-tasks are hard to optimize sometimes. But we do see the angle is hard to learn for the network.

That's awesome! The loss weight of multi-task can significantly affect the results! After changed the weight of angle_loss, I got 0.68mAP! Thank you very much!

My pleasure.

from bbavectors-oriented-object-detection.

i also want to know the loss weight of multi-task of 0.68mAP, can you share it? thanks very much! @Tarazed

from bbavectors-oriented-object-detection.

i also want to know the loss weight of multi-task of 0.68mAP, can you share it? thanks very much! @Tarazed

just the same as what the author replied, loss = hm_loss + wh_loss + off_loss + 0.1*angle_loss

from bbavectors-oriented-object-detection.

@Tarazed 请问用原始的centernet加一个额外的角度，除了main.py，还需要修改哪里嘛

from bbavectors-oriented-object-detection.

@Tarazed 请问用原始的centernet加一个额外的角度，除了main.py，还需要修改哪里嘛

Change the ground-truth accordingly in base.py

from bbavectors-oriented-object-detection.