Results 175 comments of Jiaming Han

Please try this https://github.com/csuhan/ReDet/issues/1

Please check if `work_dirs/epoch_12.pth` exists.

You can download and try our pretrained one. I guess there is something wrong with your trained model. You can try to load it mamally `model = torch.load('your_model_path')` to see...

特征图维度与offset组数无关。offset组数与anchor数量相同,即H*W个。

也有可能是container crane没有检测出object

> 我想问一下这个公式的原理是啥,为啥能从原本初始化的一些anchor中计算出具体偏移量,然后锚盒就变成高质量了? ARN类似于RPN,对preset的anchor进行refine,然后我们根据refined anchor对conv feat进行align(重采样)。 > 将原图groud truth框中的特征以包围盒方向为标准重新排列特征像素? 这里不是将conv feat于gt对齐,而是将conv feat与refined anchor(from ARN)对齐。 > ODM部分方向敏感特征和方向不变特征具体意义是什么。。 orientation-sensitive: 来自N个方向特征,对方向敏感,用于回归 orientation-invariant: 对N个方向的特征进行pooling,对方向不敏感,用于分类。

建议打印中间过程,定位卡在哪一个步骤

Yes. You can remove it. This is an experimental setting.

https://github.com/open-mmlab/mmdetection/issues/2430 may help you.