search

chenyuntc / simple-faster-rcnn-pytorch

A simplified implemention of Faster R-CNN that replicate performance from origin paper
Other
3.99k stars 1.14k forks source link

Two questions about "bbox_tools.py" #121

Open Williamongh opened 5 years ago

Williamongh commented 5 years ago

Thanks for your sharing! It helps me a lot. And I have 2 question about "./model/utils/bbox_tools.py".

  1. In "./model/utils/bbox_tools.py": 

    import numpy as np
import numpy as xp
...

    I wonder why import numpy twice as different name?

  2. In function generate_anchor_base():

    py = base_size / 2.
px = base_size / 2.

    Why set the centre of anchors (base_size / 2, base_size / 2)? Isn't (0, 0) more intuitive?

JacobYuan7 commented 5 years ago

For question2, according to the original paper, I think the author set the anchor point in the center. From the perspective of mine, (0, 0) is also okay because it can propose almost same number of anchor boxes. But what influence it will have on the mAP still needs to be tested.

Williamongh commented 5 years ago

@JacobYuan7 Thank you for your reply. I agree. For Q1, maybe it's just temporary code when debugging.

lfxx commented 4 years ago

@JacobYuan7 Thank you for your reply. I agree. For Q1, maybe it's just temporary code when debugging.

Have you tested its influence on mAP about Q2?anyway,(0,0) is more intuitive.