Awesome
<h3 id="1">Multi-shot Re-identification Based on Reinforcement Learning</h3>Training and testing codes for multi-shot Re-Identification. Currently, these codes are tested on the PRID-2011 dataset, iLiDS-VID dataset and MARS dataset. For algorithm details and experiment results, please refer our paper: Multi-shot Pedestrian Re-identification via Sequential Decision Making
<h3 id="1.1">Preparations</h3>Before starting running this code, you should make the following preparations:
- Download the MARS , iLIDS-VID and PRID-2011.
- Install MXNet following the instructions and install the python interface. Currently the repo is tested on commit e06c55.
- Download the datasets and unzip.
- Prepare data file. Generate image list file according to the file
preprocess_ilds_image.py
,preprocess_prid_image.py
andpreprocess_mars_image.py
underbaseline
folder. - The code is split to two stage, the first stage is a image based re-id task,
please refer the script
run.sh
inbaseline
folder. The codes for this stage is based on this repo. The usage is:
sh run.sh $gpu $dataset $network $recfloder
e.g. If you want to train MARS dataset on gpu 0 using inception-bn, please run:
sh run.sh 0 MARS inception-bn /data3/matt/MARS/recs
- The second stage is a multi-shot re-id task based on reinforcement learning.
Please refer the script
run.sh
inRL
folder. The usage is:
sh run.sh $gpu $unsure-penalty $dataset $network $recfloder
- For evaluation, please use
baseline/baseline_test.py
andRL/find_eg.py
. InRL/find_eg.py
, we also show some example episodes with good quality generated by our algorithm.