search

vacancy / PreciseRoIPooling

Precise RoI Pooling with coordinate gradient support, proposed in the paper "Acquisition of Localization Confidence for Accurate Object Detection" (https://arxiv.org/abs/1807.11590).
MIT License
770 stars 152 forks source link

Questions about Omega_train and loss of iou branch #45

Open zhoulukuan opened 4 years ago

zhoulukuan commented 4 years ago

Hi, i have other two questions and hope to get your help:

  1. What's the relationship between Omega_train(0.5) and normalized iou labels range([-1, 1])? In another issue #5 , I was told that iou branch labels can be normalized and achieved by 2 * (iou-0.5). I found, after filtering by the threshold of 0.5, the label of the samples is between [0.5,1]. So the normalized target labels is between [0,1]. Is this a intentionally design? If we change the Omega_train, do we need to change the normalized formula?
  2. How to compute the loss of iou branch? Dose the output of iou branch need to be activated by sigmoid layer or take an absolute value? What's the weight of the iou branch and original reg/cls branch?