Image Retrival based on VGGNet16 model application and cosine similarity.
For Linux or MacOS
- create conda virtual environment
conda create -n ${your_env_name} python=3.6
- install libraries in requirements.txt
pip install -r requirements.txt
-
Speed
-
Accuracy on big dataset, challenge scenario and challenge card
Challenge type | Speed | Accuracy |
---|---|---|
100000 templates | ss | first:% second:% |
Crop half | ss | first:% second:% |
Crop 1/3 with angle | ss | first:% second:% |
-90 Orientation | ss | first:% second:% |
random Orientation | ss | first:% second:% |
- Best dimension choice for retriavel task
- CSV
- CSV.gzip
- Pickle
- HDF5
- HDF5 zips
Conclusion
-
Pickle is the fastest in read and write, but not usable for big data which causes SystemError.
-
HDF5 is second and easy for data in structure, and also showing a good performance in compressibility.
-
Feather-format could be faster than HDF5 and Pickle.
- file size: du -h --max-depth=1 ${filename}
- file num: ls -l|grep "^-"| wc -l