re-training the model about logohunter

[ilmonteux]

For re-training the feature extractor, you will have to train the model on a logo dataset, instead of using the one pre-trained on ImageNet that comes with keras. Instead of retraining the whole model, you can probably just fine-tune it (by freezing most layers up to the very end, for example see this guide (the fine-tuning section)).
If you got the logos from the LogosInTheWild authors and followed the instructions in the data/ folder, after running src/create_clean_dataset.py you have a data/LogosInTheWild-v2/data_cleaned/brandROIs folder that contains truth-level logos extracted from the features, which should be great for fine-tuning the feature extractor. The job there would be to associate each logo to its class, and the last layers of the model would learn more logo-specific combinations of high-level features from the frozen layers. This should help with the similarity comparison (which is just the simplest decision function).

Note that re-training even the last layers of a huge model like Inception might be very hard to do even on a GPU machine. You could try a simpler model like a VGG16 first (check out load_extractor_model and model_flavor_from_name in src/utils.py for models and different truncations I tried, those can be called via the --features flag in logohunter.py).

[ilmonteux]

That would only retrain the YOLO model. You are correct that retraining the inception model on more relevant data (instead of ImageNet) would likely increase the performance for logo similarity. For example, it could be fine-tuned on a dataset made by taking the logo cut-outs from the Logosinthewild dataset (the data_cleaned/brandROIs folders mentioned here). The features extracted at the end of the encoder layers would then be more relevant to distinguish logos, so the similarity step should achieve better performance

[earlng]

I see, right, so the model you use for the Inception model is inception_logo_features.hdf5, which we download from your S3 bucket as part of the build.sh. I looked through your repo, but couldn't find any instruction on how to train an inception model. Do you have any notes for that? or point me in the right direction?

[earlng]

I see, thank you so much for the advice!