| name | caffemodel | caffemodel_url | license | caffe_commit | gist_id |
|---|---|---|---|---|---|
GoogLeNet_cars |
googlenet_finetune_web_car_iter_10000.caffemodel |
non-commercial |
737ea5e936821b5c69f9c3952d72693ae5843370 |
b90eb88e31cd745525ae |
##Description
This is the GoogLeNet model pre-trained on ImageNet classification task and fine-tuned on 431 car models in CompCars dataset. It is described in the technical report. The correspondence between the output nodes of the network and the 431 car models can be viewed at link. Please cite the following work if the model is useful for you.
A Large-Scale Car Dataset for Fine-Grained Categorization and Verification
L. Yang, P. Luo, C. C. Loy, X. Tang, arXiv:1506.08959, 2015
The bundled model is the iteration 10,000 snapshot. This model obtains a top-1 accuracy 91.2% and a top-5 accuracy 98.1% on the testing set, using only the center crop.
First, you need to download our CompCars dataset. To download, first you need to fill out an agreement according to http://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/index.html.
You will also find our train/test split files for vehicle classification on http://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/index.html. Reformulate it to any format that Caffe can read (image and label list, lmdb etc).
Then you can use the prototxt files and the model here to train, test, and extract features with the data. It uses the standard mean file of ImageNet data which can be downloaded from Caffe (try running /data/ilsvrc12/get_ilsvrc_aux.sh in Caffe).
This model uses an old version of Caffe: https://github.com/BVLC/caffe/releases/tag/rc. Please take care.
The data used to train this model comes from the ImageNet project and the CompCars dataset, which distribute their databases to researchers who agree to a following term of access: "Researcher shall use the Database only for non-commercial research and educational purposes." Accordingly, this model is distributed under a non-commercial license.