search

CharlesShang / TFFRCNN

FastER RCNN built on tensorflow
MIT License
874 stars 418 forks source link

A question about anchors #140

Open frankTian92 opened 6 years ago

frankTian92 commented 6 years ago

Sorry to bother you. As a beginner, I have been confused by the setting of numbers of anchors. In cfg, IS_MULTISCALE = False, and __C.TRAIN.ASPECTS = (1, 0.75, 0.5, 0.25) was commented. Why choosing A = len() * 3 as the numbers of anchors in lib/networks/VGGnet_train.py in line 51-54?

Benzlxs commented 6 years ago

Actually, author didn't use scale parameters in config.py like _C.TRAIN.ASPECTS and _C.TRAIN.SCALE.

The method that author employ to change the number and scale of anchor is pretty hard-coding. You can check the code below https://github.com/CharlesShang/TFFRCNN/blob/27e4a78d4ed363c7dad42fd2c140746c7425cfc8/lib/rpn_msr/anchor_target_layer.py#L33 https://github.com/CharlesShang/TFFRCNN/blob/27e4a78d4ed363c7dad42fd2c140746c7425cfc8/lib/rpn_msr/generate_anchors.py#L37

A = len() * 3 is also little bit hard-code, 3 is fixed, but it should depend on scale ratio