Evaluate with SUNRGBD Dataset with the weight file which is trained on Scenenet Dataset

[suesuekkim]

I have some questions about your model ESANet.

I trained your ESANet model on the scenenetrgbd dataset. And I want to evaluate the weight file (which calculated through scenenetrgbd dataset) in the SUNRGBD dataset. But the scenenet dataset has only 13 classes, while SUNRGBD dataset has 37 classes. Due to the number of classes are mismatched, the error occured. Is it possible to evaluate the scenenet pretrained weight file to SUNRGBD dataset?

Thank you for reading my issue.

[danielS91]

SceneNetRGBD is a synthetic dataset quite far away from reality in terms of physics and scene layout. IMHO it is only useful for pretraining in order to initialize the weights in better way than random weights would do. It does not really make sense to evaluate on SUNRGBD after training on SceneNetRGBD solely.

However, SceneNetRGBD follows the NYUv2 class spectrum with 13 classes, some of the classes match classes in the 40 (= 37 + 3 otherclasses, while the latter ones are mapped to void in SUNRGBD) class spectrum of NYUv2. You can create a simple mapping based on the class names. Moreover, you can have a look on the forward mapping from 40 to 13 classes during dataset preparation HERE.
Nevertheless, the remaining classes are missing and cannot be predicted in your setting.