/OpenVehicleReID

strong Vehicle Re-ID baseline framework.

Primary LanguagePython

open-VehicleReID

Introduction

This is a repository for vehicle Re-ID.

Update

This repo keeps updating.

2020-4-22: append apex support for saving GPU memory.

Installation

Requirements

  • Linux
  • CUDA 8.0 or higher
  • Python3
  • Pytorch 1.1+

Install open-VehicleReID

  1. Clone the open-VehicleReID repository.
    git clone https://github.com/BravoLu/open-VehicleReID.git
  1. Install the dependencies.
    cd open-VehicleReID 
    pip -r install requirements.txt

Get Started

  1. Download the dataset (VeRi776\VehicleID\VeRi_Wild).

  2. Train

    python main.py -c <path_to_config_file> \\
                   -d <path_to_data> \\
                   --gpu <gpu_ids> \\
                   --log <path_to_log_file> \\
                   --save <path_to_save_checkpoint_file> \\
                   --seed <seed_number>
    (e.g.)
    python main.py -c configs/baseline.yml -d /home/share/zhihui/VeRi/ --log logs/veri/baseline --save ckpts/veri/baseline --gpu 0,1 --seed 0
  1. Test
    python main.py -c <path_to_config_file> \\
                   -d <path_to_data> \\
                   --ckpt <path_to_load_checkpoint_file> \\
                   --gpu <gpu_ids> \\
                   --seed <seed_number> \\
                   -t 
    (e.g.)
    python main.py -c configs/baseline.yml -d /home/share/zhihui/VeRi/ --ckpt ckpts/baseline/checkpoint.pth --gpu 0,1 --seed 0

Visualization

  • We visualize the ranking list in Windows 10 (Now only support VeRi, will support other dataset in the future).
  1. Generate the ranking list pickle file.
    python main.py -c <path_to_config_file> \\
                   -d <path_to_data>  \\
                   --ckpt <path_to_load_checkpoint_file> \\
                   --gpu <gpu_ids> \\
                   --seed <seed_number>  \\
                   --vis <path_to_save_visualization_file>  \\ 
                   -t
  1. Visualize the ranking list according to the .pkl file generated in step 1.
    cd visualization
    python visualization.py --data <path_to_data>

pic

Input the file name (without extension) on the entry of Rank File(.pkl).

pic

  • (Green denotes same ids, red denotes different ids.)

More details see https://github.com/BravoLu/reid-visual-system

Benchmark

  • VehicleID
Model scale Small Median Large
Rank 1 Rank 5 Rank 10 Rank 1 Rank 5 Rank 10 Rank 1 Rank 5 Rank 10
Basline 224x224 76.4% 89.2% 92.9% 74.5% 86.5% 90.2% 71.5% 83.6% 87.7%
Basline(RR) 224x224 77.3% 89.4% 93.4% 74.5% 86.3$ 90.7% 72.2% 83.6% 87.8%
Basline 384x384 78.9% 91.3% 94.9% 76.5% 89.0% 92.9% 74.0% 86.1% 90.4%
Basline(RR) 384x384 78.9% 91.6% 95.3% 76.3% 88.4% 92.8% 74.4% 85.7% 90.0%
  • VeRi776
Model scale mAP Rank 1 Rank 5 Rank 10
Baseline 224x224 74.7% 94.8% 98.3% 99.2%
Baseline(RR) 224x224 79.8% 95.7% 97.1% 98.4%
Baseline 384x384 79.5% 95.9% 98.5% 99.0%
Baseline(RR) 384x384 84.2% 96.7% 98.4% 98.9%
  • VeRi_Wild
Model scale Small Median Large
mAP Rank1 Rank5 Rank10 mAP Rank1 Rank5 Rank10 mAP Rank1 Rank5 Rank10
Baseline 224x224 66.7% 80.6% 94.0% 97.3% 58.3% 73.0% 90.0% 95.0% 47.6% 64.8% 84.4% 90.8%
Baseline(RR) 224x224 0 0 0 0 0 0 0 0 0 0 0 0
Baseline 384x384 0 0 0 0 0 0 0 0 0 0 0 0
Baseline(RR) 384x384 0 0 0 0 0 0 0 0 0 0 0 0

* RR denotes reranking.

Acknowledgement