Closed KnightOfTheMoonlight closed 8 years ago
Notice that all the published reference models are VGG16 models. The presentation in the paper refers to the VGG16 models (even though the naming might resemble AlexNet).
A bunch of people tried to use AlexNet instead. I never tried to reproduce the experiments with the AlexNet architectures, but the folks that did found that results weren't as good. This might be due to the fact that MIL used in R*CNN relies on strong feature representations and a shallower network might be sufficient.
Hi, dear gkioxari, your network setting is based on alexnet, however you used vgg16 as your reference_model. Have you tried fine-tuning some caffemodel of alexnet?