GFocalV2
GFocalV2 copied to clipboard
implus
Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection, CVPR2021
关于训练
你好,翔哥,首先感谢你的工作.在我训练的过程中,出现了这样的问题: 我分别训练GFLv1 和 GFLv2, 发现刚开始的分类损失相差很大, GFLv1的分类损失在0.4~0.5左右,但是GFLv2的分类损失直接变成了700-800之间,我想问下这是一个正常的现象吗? 你在训练的时候遇到过吗
Since there is softmax-version focal loss, is there also a softmax version of general focal loss?
Why GIoULoss still used in GFL? To my understanding, GFL contains QFL and DFL. QFL for joint classification and IoU score while DFL for bbox regression  So DFL is...
你好，我在单目标跟踪中发现回归分支收敛很慢，且模型的精度很差，DFL基本在0.9左右，epoch调整到50后，DFL能到0.75，但是精度和epoch为20时候差不多，并且在测试阶段，平均的IOU很低，我想指导在检测上DFL以及GIOU loss大概能到多少
既然QFocalLoss 和Distribution_focal_loss 两者相互带来增益，那是否可以表示为GFocal = QFocalLoss + Distribution_focal_loss
Does correlation between real localization quality and top-1 value of general distribution are consistent during training? I'm wonder that those relations are consistent during whole training time not just after...
@implus 如题，按照我的理解，我将训练集1个epoch的regression_target保存了下来（shuffle设定为False），并可视化结果如下图所示。  感觉和论文中翔哥你展示的不一致。  我的做法是在gfl_head.py找那个将target_corners这个变量保存下来，然后画直方图，请问有什么问题吗？ ``` with open('./target_corners_results/{}.npy'.format(time.time()), 'wb') as f: np.save(f, target_corners.cpu().numpy()) ```
hellow ，When I tried to run test.py with vscode, an error occurred: KeyError: 'Gfocalhead is not in the head registry', The command bash./tools/dist_test.sh configs/gfocal/gfocal_r50_fpn_ms2x.py work_dirs/gfocal_r50_fpn_ms2x/latest.pth 1 to run the dist_test.sh...
