Thinklab-SJTU / R3Det_Tensorflow

Code for AAAI 2021 paper: R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object
Apache License 2.0
546 stars 122 forks source link

Adapting Anchors & Angles for Extremely Skewed Dataset #121

Closed dilpreetsingh closed 3 years ago

dilpreetsingh commented 3 years ago

Hey,

My dataset has extremely skewed aspect ratios, so for example the width of an object could be 800px, whilst the height may only be 20px. So the objects are extremely wide and thin, and generally run across from left to right of the image.

I was wondering whether/if I need to adapt the ANCHOR_RATIOS and ANCHOR_ANGLES (in libs/configs/cfgs.py) to my own dataset? Or should I use the default ratios?

Any other recommendations/tips for dealing with such skewed objects?

yangxue0827 commented 3 years ago

anchor free based method may be more suitable