Which pretrained model is better if I finetune it on new dataset？

[catsdogone]

I finetune the pvanet/full/test.model using the train.prototxt in pvanet/example_fineturn on KITTI while the result is not good and even worse than the pvanet/full/test.model. Should I use imagenet/full/test.model as pre-train model if I want to train on a dataset with different classes with VOC? Can you please give me some advice? Thank you very much. #27

[sanghoon]

Sorry for a late comment. I'm attending a conference right now, I'll check the issue when I get back to my home.

However, I guess it may be a training parameter issue, please check #21 Currently provided prototxts have 'rpn_pre_nms_top_k = 200~300' even in the training phases, but it shouldn't be (it's our fault). It must be around 2000 in the training phase

[catsdogone]

@sanghoon Thank you. I have corrected the rpn_pre_nms_top_k from 200 to 2000 in the prototxt. But the result still not good enough as describe in #27 . The attachment is my solver and script file. I am still trying to figure out the reason. kitti_pva.txt solver_plateau.txt train.txt

[songjmcn]

I think because the pva-net not good for small object。I think you will redefine the network structure

[PaTricksStar]

@catsdogone Hi catsdogone. Do you get better results on KITTI? It seems that PVANet should work better than the original Faster RCNN on KITTI for the more complicate network and multi-layer features. Have you tried the ImageNet pretrained models?

[scofield1991]

So guys, was anybody able to train PVA on a KITTI dataset?