Transfer learning performance about squeezedet HOT 3 CLOSED

@amrit110 Well, there can be a lot of reasons. If you could point me to your repository and share more details of your training procedure, that's going to be helpful for us to locate the problems.

from squeezedet.

sorry false alarm. Ok so, i used a model trained on KITTI but changed the config file to use demo.py on cityscapes images. When I said overfitting, that was not really what i meant because a lightweight model like yours should be even less prone to overfitting. I was concerned that the model performed well only on KITTI images (camera config and conditions). So, then i realised that the image size i was using from cityscapes was the original (2048 x 1024) which is why i got bad performance. I rescaled it to something closer to the training set (KITTI) and then i get expected performance. Now, I am trying to actually train it on cityscapes. great work, love the low inference time. if i manage to train on cityscapes successfully, ill share the results.

from squeezedet.

@amrit110 Great, looking forward to your cityscape result.

from squeezedet.