Tips for improving search result? retraining?
dgtlmoon opened this issue · comments
Thanks for the amazing work!
Given I have a fairly specific domain I'm searching, for example always print artwork on tshirts, or pictures of cow patterns.. or whatever, what steps would you recommend to tune the network? some things I thought of for the actual image which helped my results are to reduce noise, trim image automatically to only include the relevant data and apply contrast/sharpness/other general noise removal algorithms..
I've also tried changing the layer (fc1
versus fc2
) and tweaking the PCA number of components.
Would you recommend a different network that's supported by keras, like InceptionResNetV2
etc?
But what about tweaking the weights of the actual network vgg16_weights_tf_dim_ordering_tf_kernels.h5
? would it be worth retraining those weights on my fairly huge and domain-specific image set (200,000+ images)? can you recommend any reading material? what's your thoughts?
How about changing the similarity grouping? I found that cosine worked best for me, euclidean not so much
So I guess to sum it up, given one has a very domain specific image set they are searching, what are some ways to improve the results?