Relja
commented
5 years ago Hi,
Thanks for your interest in our work and good luck with your project.
Q1) Yes
Q2) That sounds perfectly reasonable and is how I would have approached it.
As I understand it all weights belonging to the VGG part are fixed and only the NetVlad layer is trained on the dataset.
No, the backPropToLayer option is governing which parts of the network are trained, i.e. from conv5_3 onwards it is trained. So you're training conv5_3 + NetVLAD. The reason I stopped at conv5_3 is purely due to GPU RAM, if you can backprop further (and are not afraid of overfitting) then you can go deeper.
Regarding the error - I never saw that and can't really see there being a bug in that piece of code as it is quite simple. It goes through each query, gets its potential positives and computes distances between the two. So presumably potential positives are wrong, there is an ID there that is larger than the total number of images. You can easily check them by looping through queries and examining the max of db.nontrivialPosQ(qID) .
You don't need to rerun training to get to the same error - the files are all there so you should be able to just load the features and rerun testCoreRank. Could be useful to load the files and see their sizes, if they're what you expect. Presumably dbFeat's second dimension is 33935, is this the size of your dataset?
Amadeeus
Dear Relja, In the context of my Master's thesis I am working on visual place recognition and I am using NetVlad as a baseline. In order to make the comparison a bit more fair, I would like to fine-tune your best performing (compact-descriptor) model on my dataset, which is a subset of the Oxford Robotcar dataset.
I have two questions and one issue that I would very much appreciate your help with: Question 1) For using NetVlad as baseline, I want to fine-tune your best performing compact-descriptor model, which according to my understanding would be that one contained in _vd16_pitts30k_conv5_3_vlad_preL2_intrawhite plus cropping the resulting descriptors at 256-Dims and L2-normalizing them. Could you just quickly confirm this?
Question 2) Is there an easy way to adjust the code such that I can take exactly that model downloaded from your website and train it further for a couple of epochs on my dataset? What I tried so far is summarized as follows: 1) I defined the dataset exactly as described in
dbBase.m. 2) I am 100% sure that there are positives in all the train / val / test sets. 3) I modiefied lines 4-26 inloadNet.mto be able to load the model downloaded from your project website:4) My train_robotcar.m looks like this:
As I understand it all weights belonging to the VGG part are fixed and only the NetVlad layer is trained on the dataset. Here is the list of options output when I start the training:
After training for a couple of hours I get the following error:
Can you make anything of this? Your help is much appreciated.