ETR: An Efficient Transformer for Re-ranking in Visual Place Recognition
Feature Extraction
Please follow the instruction to install the DELG library.
All the instructions below assume the DELG package is installed in DELG_ROOT.
please first read the instrcution, which is writed for the feature extraction of Revisited Oxford/Paris. And make sure that you fully understand the instrcution.
Next, we take the feature extraction of Tokyo247 as an example to explain the process of the DELG feature extraction.
Extract the features of Tokyo247
Download the test set of Tokyo247 first, then modify the path of the images in the file dataset/Tokyo247/tokyo247_db_c.txt and dataset/Tokyo247/tokyo247_query_c.txt.
Then we need copy-paste the file delg_feature_extract.py and delg_feature_extract.sh to the DELG_ROOT/delf/python/delg. use the shell script delg_feature_extract.sh to extract features. Note that the scriptes may not work out-of-the-box, you may still need to set the paths of the input/output directories properly.
Evaluation
ETR Reranking
Before reranking, we need to generate a initial rank result. Take Tokyo247 as an example, the initial rank file is dataset/Tokyo247/delg_feats/delg_tokyo247_rank_index.npy. you can generate a initial rank file by yourself use any other global retrieval methods. Then modify the python script evaluate_etr.py, change the dataset_name to dataset that you want to evaluate and run this:
python evaluate_etr.py