Could you please give the detailed information about Ensemble? about stcn HOT 4 CLOSED

Without using BL30K data, I tried to use resnet101+aspp, resnet101w+aspp, and resnet50w+aspp, the performance on YTB all dropped lightly. I am not quite sure of whether I need to tune the hypers more carefully for better performance. Have you used BL30K data? @longmalongma

from stcn.

Without using BL30K data, I tried to use resnet101+aspp, resnet101w+aspp, and resnet50w+aspp, the performance on YTB all dropped lightly. I am not quite sure of whether I need to tune the hypers more carefully for better performance. Have you used BL30K data? @longmalongma

@JerryX1110 Yes, I tried to use resnet101+aspp, resnet101w+aspp, and resnet50w+aspp with BL30K, the performance on YTB all also dropped lightly. That's why I brought it up.

from stcn.

@JerryX1110 @longmalongma Sorry that I could not be more responsive these days since I am moving and there is a lot to handle.

From my experience changing the backbone from resnet50 to wide-resnet50 always helps. Adding ASPP or using a deeper network does not always help because it changes the receptive field and you might need a larger crop size in training.

from stcn.

Just to clarify -- I did not use a larger crop size in my experiments for models with ASPP because it's so much slower to train. That's just a suggestion.

from stcn.