Related Project: Strong Identification Loss Baseline
This project provides a strong triplet loss baseline in person re-identification, using pytorch.
Triplet loss with settings:
- ResNet-50,
stride = 2orstride = 1in last conv block - NOT normalizing feature to unit length, with margin 0.3
- Only horizontal flipping used for data augmentation
im_w x im_h = 128 x 256,ims_per_id = 4,ids_per_batch = 32- Adam optimizer, base learning rate 2e-4, decaying exponentially after 150 epochs. Train for 300 epochs in total.
The results are as follows. S1 and S2 means stride = 1 and stride = 2 respectively; R.R. means using re-ranking.
| Rank-1 (%) | mAP (%) | R.R. Rank-1 (%) | R.R. mAP (%) | |
|---|---|---|---|---|
| Market1501-S2 | 86.43 | 71.50 | 89.82 | 85.55 |
| Market1501-S1 | 89.04 | 75.29 | 91.57 | 87.82 |
| Duke-S2 | 78.82 | 61.09 | 83.98 | 79.15 |
| Duke-S1 | 79.76 | 64.27 | 85.32 | 81.48 |
| CUHK03-S2 | 56.36 | 50.82 | 65.21 | 65.76 |
| CUHK03-S1 | 59.14 | 54.43 | 69.86 | 70.03 |
We see that stride = 1 (higher spatial resolution before global pooling) has obvious improvement over stride = 2 (original ResNet). I tried this inspired by paper Beyond Part Models: Person Retrieval with Refined Part Pooling.
Other details of setting can be found in the code. To test my trained models or reproduce these results, see the Examples section.
This repository contains following resources
- A beginner-level dataset interface independent of Pytorch, Tensorflow, etc, supporting multi-thread prefetching (README file is under way)
- Three most used ReID datasets, Market1501, CUHK03 (new protocol) and DukeMTMC-reID
- Python version ReID evaluation code (Originally from open-reid)
- Python version Re-ranking (Originally from re_ranking)
- Triplet Loss
It's recommended that you create and enter a python virtual environment, if versions of the packages required here conflict with yours.
I use Python 2.7 and Pytorch 0.3. For installing Pytorch, follow the official guide. Other packages are specified in requirements.txt.
pip install -r requirements.txtThen clone the repository:
git clone https://github.com/huanghoujing/person-reid-triplet-loss-baseline.git
cd person-reid-triplet-loss-baselineInspired by Tong Xiao's open-reid project, dataset directories are refactored to support a unified dataset interface.
Transformed dataset has following features
- All used images, including training and testing images, are inside the same folder named
images - Images are renamed, with the name mapping from original images to new ones provided in a file named
ori_to_new_im_name.pkl. The mapping may be needed in some cases. - The train/val/test partitions are recorded in a file named
partitions.pklwhich is a dict with the following keys'trainval_im_names''trainval_ids2labels''train_im_names''train_ids2labels''val_im_names''val_marks''test_im_names''test_marks'
- Validation set consists of 100 persons (configurable during transforming dataset) unseen in training set, and validation follows the same ranking protocol of testing.
- Each val or test image is accompanied by a mark denoting whether it is from
- query (
mark == 0), or - gallery (
mark == 1), or - multi query (
mark == 2) set
- query (
You can download what I have transformed for the project from Google Drive or BaiduYun. Otherwise, you can download the original dataset and transform it using my script, described below.
Download the Market1501 dataset from here. Run the following script to transform the dataset, replacing the paths with yours.
python script/dataset/transform_market1501.py \
--zip_file ~/Dataset/market1501/Market-1501-v15.09.15.zip \
--save_dir ~/Dataset/market1501We follow the new training/testing protocol proposed in paper
@article{zhong2017re,
title={Re-ranking Person Re-identification with k-reciprocal Encoding},
author={Zhong, Zhun and Zheng, Liang and Cao, Donglin and Li, Shaozi},
booktitle={CVPR},
year={2017}
}
Details of the new protocol can be found here.
You can download what I have transformed for the project from Google Drive or BaiduYun. Otherwise, you can download the original dataset and transform it using my script, described below.
Download the CUHK03 dataset from here. Then download the training/testing partition file from Google Drive or BaiduYun. This partition file specifies which images are in training, query or gallery set. Finally run the following script to transform the dataset, replacing the paths with yours.
python script/dataset/transform_cuhk03.py \
--zip_file ~/Dataset/cuhk03/cuhk03_release.zip \
--train_test_partition_file ~/Dataset/cuhk03/re_ranking_train_test_split.pkl \
--save_dir ~/Dataset/cuhk03You can download what I have transformed for the project from Google Drive or BaiduYun. Otherwise, you can download the original dataset and transform it using my script, described below.
Download the DukeMTMC-reID dataset from here. Run the following script to transform the dataset, replacing the paths with yours.
python script/dataset/transform_duke.py \
--zip_file ~/Dataset/duke/DukeMTMC-reID.zip \
--save_dir ~/Dataset/dukeLarger training set tends to benefit deep learning models, so I combine trainval set of three datasets Market1501, CUHK03 and DukeMTMC-reID. After training on the combined trainval set, the model can be tested on three test sets as usual.
Transform three separate datasets as introduced above if you have not done it.
For the trainval set, you can download what I have transformed from Google Drive or BaiduYun. Otherwise, you can run the following script to combine the trainval sets, replacing the paths with yours.
python script/dataset/combine_trainval_sets.py \
--market1501_im_dir ~/Dataset/market1501/images \
--market1501_partition_file ~/Dataset/market1501/partitions.pkl \
--cuhk03_im_dir ~/Dataset/cuhk03/detected/images \
--cuhk03_partition_file ~/Dataset/cuhk03/detected/partitions.pkl \
--duke_im_dir ~/Dataset/duke/images \
--duke_partition_file ~/Dataset/duke/partitions.pkl \
--save_dir ~/Dataset/market1501_cuhk03_dukeThe project requires you to configure the dataset paths. In tri_loss/dataset/__init__.py, modify the following snippet according to your saving paths used in preparing datasets.
# In file tri_loss/dataset/__init__.py
########################################
# Specify Directory and Partition File #
########################################
if name == 'market1501':
im_dir = ospeu('~/Dataset/market1501/images')
partition_file = ospeu('~/Dataset/market1501/partitions.pkl')
elif name == 'cuhk03':
im_type = ['detected', 'labeled'][0]
im_dir = ospeu(ospj('~/Dataset/cuhk03', im_type, 'images'))
partition_file = ospeu(ospj('~/Dataset/cuhk03', im_type, 'partitions.pkl'))
elif name == 'duke':
im_dir = ospeu('~/Dataset/duke/images')
partition_file = ospeu('~/Dataset/duke/partitions.pkl')
elif name == 'combined':
assert part in ['trainval'], \
"Only trainval part of the combined dataset is available now."
im_dir = ospeu('~/Dataset/market1501_cuhk03_duke/trainval_images')
partition_file = ospeu('~/Dataset/market1501_cuhk03_duke/partitions.pkl')Datasets used in this project all follow the standard evaluation protocol of Market1501, using CMC and mAP metric. According to open-reid, the setting of CMC is as follows
# In file tri_loss/dataset/__init__.py
cmc_kwargs = dict(separate_camera_set=False,
single_gallery_shot=False,
first_match_break=True)To play with different CMC options, you can modify it accordingly.
# In open-reid's reid/evaluators.py
# Compute all kinds of CMC scores
cmc_configs = {
'allshots': dict(separate_camera_set=False,
single_gallery_shot=False,
first_match_break=False),
'cuhk03': dict(separate_camera_set=True,
single_gallery_shot=True,
first_match_break=False),
'market1501': dict(separate_camera_set=False,
single_gallery_shot=False,
first_match_break=True)}You can use a trained model to extract features for a list of images, and then perform whatever you desire with these features. An example is
python script/experiment/infer_images_example.py \
--model_weight_file YOUR_MODEL_WEIGHT_FILEMy training log and saved model weights for three datasets can be downloaded from Google Drive or BaiduYun.
Specify
- a dataset name (one of
market1501,cuhk03,duke) - stride,
1or2 - an experiment directory for saving testing log
- the path of the downloaded
model_weight.pth
in the following command and run it.
python script/experiment/train.py \
-d '(0,)' \
--only_test true \
--dataset DATASET_NAME \
--last_conv_stride STRIDE \
--normalize_feature false \
--exp_dir EXPERIMENT_DIRECTORY \
--model_weight_file THE_DOWNLOADED_MODEL_WEIGHT_FILEYou can also train it by yourself. The following command performs training, validation and finally testing automatically.
Specify
- a dataset name (one of
['market1501', 'cuhk03', 'duke']) - stride,
1or2 - training on
trainvalset ortrainset (for tuning parameters) - an experiment directory for saving training log
in the following command and run it.
python script/experiment/train.py \
-d '(0,)' \
--only_test false \
--dataset DATASET_NAME \
--last_conv_stride STRIDE \
--normalize_feature false \
--trainset_part TRAINVAL_OR_TRAIN \
--exp_dir EXPERIMENT_DIRECTORY \
--steps_per_log 10 \
--epochs_per_val 5During training, you can run the TensorBoard and access port 6006 to watch the loss curves etc. E.g.
# Modify the path for `--logdir` accordingly.
tensorboard --logdir YOUR_EXPERIMENT_DIRECTORY/tensorboardFor more usage of TensorBoard, see the website and the help:
tensorboard --helpSpecify
- a dataset name (one of
['market1501', 'cuhk03', 'duke']) - stride,
1or2 - either
model_weight_file(the downloadedmodel_weight.pth) ORckpt_file(savedckpt.pthduring training) - an experiment directory for saving images and log
in the following command and run it.
python script/experiment/visualize_rank_list.py \
-d '(0,)' \
--num_queries 16 \
--rank_list_size 10 \
--dataset DATASET_NAME \
--last_conv_stride STRIDE \
--normalize_feature false \
--exp_dir EXPERIMENT_DIRECTORY \
--model_weight_file '' \
--ckpt_file ''Each query image and its ranking list would be saved to an image in directory EXPERIMENT_DIRECTORY/rank_lists. As shown in following examples, green boundary is added to true positive, and red to false positve.
Test with CentOS 7, Intel(R) Xeon(R) CPU E5-2618L v3 @ 2.30GHz, GeForce GTX TITAN X.
Note that the following time consumption is not gauranteed across machines, especially when the system is busy.
For following settings
- ResNet-50,
stride=1in last block identities_per_batch = 32,images_per_identity = 4,images_per_batch = 32 x 4 = 128- image size
h x w = 256 x 128
it occupies ~11000MB GPU memory.
If not having a 12 GB GPU, you have to either decrease identities_per_batch or use multiple GPUs.
Taking Market1501 as an example, it contains 31969 training images of 751 identities, thus 1 epoch = 751 / 32 = 24 iterations. Each iteration takes ~1.08s, so each epoch ~27s. Training for 300 epochs takes ~2.25 hours.
Taking Market1501 as an example
- With
images_per_batch = 32, extracting feature of whole test set (12936 images) takes ~160s. - Computing query-gallery global distance, the result is a
3368 x 15913matrix, ~2s - Computing CMC and mAP scores, ~15s
- Re-ranking requires computing query-query distance (a
3368 x 3368matrix) and gallery-gallery distance (a15913 x 15913matrix, most time-consuming), ~90s
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent