search

thangvubk / Cascade-RPN

Code for NeurIPS 2019 paper: "Cascade RPN: Delving into High-Quality Region Proposal Network with Adaptive Convolution"
Apache License 2.0
180 stars 19 forks source link

Why do you use a single anchor per location? #3

Closed bruceyang2012 closed 4 years ago

bruceyang2012 commented 5 years ago

Hi，I wonder why you use a single anchor per location. Will you get a better result if using more anchor ratios? Thanks.

zlj199502 commented 5 years ago

I want to know about it too.

thangvubk commented 5 years ago

The are two reasons (1) remove hyper-parameter defined by anchor scales and ratios and (2) reduce overhead of feature sampling for adaptive conv. When using more scales, it is expected that the performance is at least comparable to 1 scale.

zlj199502 commented 5 years ago

If I want using more anchors. How can I do the adaptive conv?

thangvubk commented 5 years ago

The feature sampling of adaptive convolution is controlled by the offsets. You need to modify anchor_offset() to support multiple anchors per location.

zlj199502 commented 5 years ago

Thank you for your reply. I mean if i get multiple offsets. How to combined with multiple anchor?

thangvubk commented 5 years ago

@zlj199502 you may need to utilize deformable group to use multiple multiple anchor.