libuyu

The target is expected to have the one-hot format. If the target is the class number from COCO, these codes are used to deal with it: https://github.com/libuyu/mmdetection/blob/be06992564cc6b995b1ae86a258568e9d7b7a599/mmdet/models/losses/ghm_loss.py#L47-L49

1&2: The exponential moving average is just a widely used technique to keep a variable more stable during updating [(see wiki)](https://en.wikipedia.org/wiki/Moving_average). The SGD optimizer also adopts it and has a...

Your understanding is right, and our target is just making the weight of these samples small. The motivation and details can be seen in our paper https://arxiv.org/abs/1811.05181

@DHPO You are right. In the paper, we define the M as the number of all bins. And in the latest version of our code, we choose the number of...

Thank you for your affirmation. It seems you still keep some configs of COCO, because the ignore region is used in COCO with the crowd label but not in VOC....

Sorry, I forgot that. I just use the official model. https://download.pytorch.org/models/resnet50-19c8e357.pth

The model is trained on image-net and in theory, it can be used for fine-tuning on any data set.

I understand your concern. But I think both COCO and ImageNet are very large compared with your set. So maybe the pre-trained model is not the key. Although you can...

You can refer to Equation 9-11 in the paper. And note that the loss is divided by the total sample number in the end: https://github.com/libuyu/mmdetection/blob/be06992564cc6b995b1ae86a258568e9d7b7a599/mmdet/models/losses/ghm_loss.py#L75-L76 So the weights here correspond...

@longchuan1985 We use the batch size of 16 (8 GPUs with two images per GPU), which can be seen in the example script. And I want to clarify that the...