Can I use other data sets for this experiment

[xychandsome]

I want to use other object detection data sets for this experiment, can I do this, such as Pascal VOC,thanks!

[ChrisAllenMing]

The DM-ADA model is specifically designed for classification tasks, while I think it can be applied to cross-domain detection tasks with some modifications:

• For data loader, you can refer the codes of our another work.
• For the domain mixup strategy, you can apply feature-level mixup to the feature map of whole image or the feature vector obtained by ROI pooling, and use these mixup features for adversarial domain adaptation.

[xychandsome]

The DM-ADA model is specifically designed for classification tasks, while I think it can be applied to cross-domain detection tasks with some modifications:

* For data loader, you can refer [the codes  of our another work](https://github.com/ChrisAllenMing/GPA-detection/tree/master/lib/datasets).

* For the domain mixup strategy, you can apply feature-level mixup to [the feature map of whole image](https://github.com/ChrisAllenMing/GPA-detection/blob/d63d3cd8a76523615c288cceb676f75967f5914e/lib/model/faster_rcnn/faster_rcnn.py#L47) or [the feature vector obtained by ROI pooling](https://github.com/ChrisAllenMing/GPA-detection/blob/d63d3cd8a76523615c288cceb676f75967f5914e/lib/model/faster_rcnn/faster_rcnn.py#L88), and use these mixup features for adversarial domain adaptation.

OK,thanks,I will try it! And I want to know the meaning of line 139 in trainer_mixer.py src_inputs_unnorm = (((src_inputs self.std[0]) + self.mean[0]) - 0.5) 2 What is this step used for?

[ChrisAllenMing]

This line of code maps an image to its unnormalized version, and this unnormalized counterpart is used to get the pixel-level mixup samples, as shown in this line of code. Also, we feed these unnormalized images to domain discriminator for adversarial training, as shown in this line of code.