Skip to content

reid baseline model for exploring softmax and triplet hard loss (Gluon implementation)

Notifications You must be signed in to change notification settings

jhyching/reid_baseline_gluon

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ReID_baseline

Baseline model (with bottleneck) for person ReID (using softmax and triplet loss).

We support

  • multi-GPU training
  • easy dataset preparation
  • end-to-end training and evaluation

Get Started

  1. cd to folder where you want to download this repo

  2. Run git clone https://github.com/L1aoXingyu/reid_baseline_gluon.git

  3. Install dependencies:

    pip install --pre mxnet-cu90
    
    • tensorflow (for tensorboard)
    • MXBoard
  4. Prepare dataset

    Create a directory to store reid datasets under this repo via

    cd reid_baseline
    mkdir data
    1. Download dataset to data/ from http://www.liangzheng.org/Project/project_reid.html
    2. Extract dataset and rename to market1501. The data structure would like:
    market1501/
        bounding_box_test/
        bounding_box_train/
    
  5. Prepare pretrained model if you don't have

    from mxnet import gluon
    gluon.model_zoo.vision.resnet50_v1(pretrained=True)

    Then it will automatically download model in ~.mxnet/models/, you should set this path in config.py

Train

You can run

bash scripts/train_triplet_softmax.sh

in reid_baseline folder if you want to train with softmax and triplet loss. You can find others train scripts in scripts.

Results

loss rank1 map
softmax 87.1% 67.8%
triplet 88.2% 73.7%
triplet + softmax 90.4% 76.4%

I find the mxnet.gluon results are a little bit lower than pytorch results, and I cannot get the reason. I would appreciate that if anyone can help me.

About

reid baseline model for exploring softmax and triplet hard loss (Gluon implementation)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.0%
  • Shell 1.0%