This project tries to build a benchmark for image retrieval, particularly for Instance-level image retrieval.
Fisher vector and VLAD are both currently working on the scores dataset. To run them,
- install
score-retrieval
, - run
make
, follow the instructions, then runmake
again, - run
make yael
, - run
make hesaff
, - run
make setup-dataset
, - run the commands in the
fv
orvlad
make target manually (make fv
/make vlad
isn't working right now for some reason).
Note: The training parameters in gmm.py
(for FV) and kmeans.py
(for VLAD) have been massively reduced for ease of testing. If you want a real training run, increase them.
method | feature | mAP (best) | status | links |
---|---|---|---|---|
fc_retrieval | CNN | 60.2% | finished | fc_retrieval |
rmac_retrieval | CNN | to be tested | finished | rmac_retrieval |
crow_retrieval | CNN | to be tested | finished | crow_retrieval |
fv_retrieval | SIFT | 67.29% | finished | fv_retrieval |
vlad_retrieval | SIFT | 63.13% | finished | vlad_retrieval |
the methods on above have the following characteristics:
- Low dimension
- Time - tested
- Used in industry
If you are interested in this project, feel free to contribute your code. Only Python and C++ code are accepted.