Skip to content

Latest commit

 

History

History
executable file
·
57 lines (40 loc) · 2.14 KB

README.md

File metadata and controls

executable file
·
57 lines (40 loc) · 2.14 KB
Raw

Budgeted Stream Simulation

Code base for simulation on the budgeted stream person re-ID for paper Resource Aware Person Re-identification across Multiple Resolutions.

This code generates Figure 6 in the paper.

Requirement

  • python
  • numpy
  • colorlog (for logging)
  • scipy (for reading .mat file)
  • matplotlib (for plotting)
  • seaborn (for plotting)

How to run

Data

Feature vectors: You should use the trained network to generate feature vectores for query and gallery images at stage 1-3 and fusion stage. We use .mat file type in this code. Name them as query_features_[1-3|fusion].mat and test_features_[1-3|fusion].mat and store them in a folder (e.g. ./data/feature/DaRe)

Original data: Original data are needed for collecting data labels. You can download Market-1501-v15.09.15 from here, and unzip it in a folder (e.g. ./dataset)

Commands

Run all

./simulation.sh <dataset_path> <feature_path>

DaRe(R)+RE (distance)

python main.py --log_file distance_confidence --confidence_function distance --dataset_path <dataset_path> --feature_path <feature_path> --test_budget

DaRe(R)+RE (margin)

python main.py --log_file margin_confidence --confidence_function margin --dataset_path <dataset_path> --feature_path <feature_path> --test_budget

DaRe(R)+RE (random)

python main.py --log_file random --confidence_function random --dataset_path <dataset_path> --feature_path <feature_path> --test_budget

Plot

python budgeted_stream_plot.py

Notes

  • Use --dump_distance_mat option to dump the resulted distance matrices and use this widely used evaluation code to evaluate the performance (the result should be the same).
  • Use --dump_exit_history option to dump exit history for each query image, and use these data to generate qualitative results in section 5.4. (We use the history at budget = 1030203460.27 (corresponding to q = 0.5) when using distance confidence function)

License

MIT