In this repository, we provide the code used for our paper A Pose-Sensitive Embedding for Person Re-Identification with Expanded Cross Neighborhood Re-Ranking.

This includes our training and prediction framework as well as the used neural network architectures and dataset readers. The Matlab code for our Expanded Cross Neighborhood Re-Ranking is located in a separate repository. This training framework is based on Google's Tensorflow and the original network architectures (Resnet and Inception-v4) are inspired by the implementations provided in the Tensorflow Models repository. All code is written in Python3 and was used with Tensorflow 1.3.

If you find our work helpful in your research, please cite:

M. Saquib Sarfraz, Arne Schumann, Andreas Eberle, Ranier Stiefelhagen,
"A Pose Sensitive Embedding for Person Re-Identification with Exapanded Cross Neighborhood Re-Ranking",
https://arxiv.org/abs/1711.10378 2017

In order to train our models for Person Re-Identification, we start of with an Imagenet pre-trained model and fine tune it on our task. These pre-trained models can be found on the Tensorflow Models Readme Page.

For training a Person Re-Id model, the trainer_preid.py script is used. To get a look at all possible arguments, have a look at the script's main method. Here, we'll only show the most important ones.

To be able to use the dataset readers, simply extract the datasets downloaded from their project websites into an empty folder.

We currently support duke, market1501 and mars. Although there are some other readers provided, we cannot give guarantees for them. To see the names used for training these datasets, have a look at the DatasetFactory.py class.

To see the supported network's and their names, best have a look at the nets_factory.py script. Please note that the Tensorflow Models framework refers to the Resnet-50 network as resnet_v1_50. For the sake of compatibility, we did not change that.

To run our training script for a Resnet-50 Baseline model with market, execute the following while replacing all the <> tags with the corresponding values.

python3 trainer_preid.py --output=<output directory> --data=<dataset directory> --dataset-name=market1501 --batch-size=16 --num-epochs=100 --network-name=resnet_v1_50 --initial-checkpoint=<path to imagenet checkpoint or another checkpoint you want to load> --checkpoint-exclude-scopes=resnet_v1_50/logits --trainable-scopes=resnet_v1_50/logits --no-evaluation

In this example, we use the optional parameters --checkpoint-exclude-scopes and --trainable-scopes. With the former, we can specify scopes to be excluded when loading the initial checkpoint (e.g. here, we exclude the logits, as they do not match between imagnet and Market1501). These layers will be randomly initialized. The latter one allows us to specify the scopes that should be trained. If this parameter is not specified, the whole network will be trained.

For our work, we always started with an Imagenet pre-trained model and first trained only the randomly initialized layers before fine-tuning the whole network. To do this, you would start a new training after the one above finished or converged. For this new training, you specify that result as initial checkpoint and do not exclude scopes and train the whole network (not specify --trainable-sopes).

To find out the scope names, please have a look at the network implementations in the nets package.

If you do not specify the --no-evaluation flag, the trainer_preid.py script will evaluate the model with the test and query set after every epoch. This is done by first predicting the test and query set and then using the Market1501 matlab evaluation script. Furthermore, the script will store the best checkpoint with the predicted features. Please note that this is not implemented for the MARS dataset as it takes a very long time predicting the features. These best features will be located in a subfolder of the output directory called predictions-best.

In order to train our Views model, you need to have a dataset providing views information. In our paper, we used RAP for training the view predictor before transfering it to one of the other datasets. In contrast, our RAP dataset does not utilize person labels and thus can only be used for views training. The trainer_views.py script can be called as follows:

python3 trainer_views.py --output=<output directory> --data=<dataset directory> --dataset-name=market1501 --batch-size=16 --num-epochs=100 --network-name=resnet_v1_50_views --initial-checkpoint=<path to imagenet checkpoint or another checkpoint you want to load> --checkpoint-exclude-scopes=resnet_v1_50/logits --trainable-scopes=resnet_v1_50/3Views

The pose maps models are basically used the same way as the Baseline and the View models. However, to use them, you first need to generate the pose maps for the images. We did this by using the Deeper Cut Tensorflow implementation. For handling the pose maps, the datasets in our framework also have a <dataset-name>-pose-maps counterpart.

Therefore, to train a pose maps Resnet-50 model, you can run the following:

python3 trainer_preid.py --output=<output directory> --data=<dataset directory> --dataset-name=market1501-pose-maps --batch-size=16 --num-epochs=100 --network-name=resnet_v1_50 --initial-checkpoint=<path to imagenet checkpoint or another checkpoint you want to load> --checkpoint-exclude-scopes=resnet_v1_50/logits,resnet_v1_50/conv1 --trainable-scopes=resnet_v1_50/logits,resnet_v1_50/conv1 --no-evaluation

Please note that we mainly change the --dataset-name parameter to market1501-pose-maps. However, as the Imagenet pre-trained model expects only three input layers, we cannot use the first layer of that model. Therefore we not only exclude the logits, but also the conv1 layer. Accordingly, we also add this layer to the trainable layers to enable the network to learn this layer in the first training step.

Our PSE models are a combination of the Views Model and the Pose Maps Model. Therefore, for training the model, the training processes are also combined. The following command can be used to train a Resnet 50 PSE model.

python3 trainer_preid.py --output=<output directory> --data=<dataset directory> --dataset-name=market1501-pose-maps --batch-size=16 --num-epochs=100 --network-name=resnet_v1_50_views --initial-checkpoint=<path to model with RAP pre-trained views predictor> --checkpoint-exclude-scopes=resnet_v1_50/logits --trainable-scopes=resnet_v1_50/logits,resnet_v1_50/pre_logits,resnet_v1_50/3ViewBranches --no-evaluation

Herem we use the dataset version providing pose maps (market1501-pose-maps) with the network architecture utilizing the views prediction (resnet_v1_50_views). It is important to note that we have to use an initial model with a trained views predictor. This is the case because we do not have view labels for the Market1501 and Duke dataset. Again, we first train all randomly initialized layers before fine-tuning the whole model afterwards.

During training, you can keep track of the loss and other important numbers by starting Tensorboard.

tensorboard --logdir=<output directory of the training or a parent folder of it>

The results can be viewed by opening a browser and go to localhost:6006.

To predict the features, run the predictor_preid.py script:

python3 predictor_preid.py --model-dir=<the model to be loaded> --data=<dataset directory> --dataset-name=market1501 --batch-size=128 --network-name=resnet_v1_50_views

The predicted features will be stored in a subfolder of the specified model-dir called predictions.

You can download our deployed PSE model's features for Duke and Market here

https://drive.google.com/drive/folders/1uyTMejyhpkPD8L-v8JJ9JOECrEIYZPCI?usp=sharing

Please see Expanded Cross Neighborhood Re-Ranking to evaluate and reproduce the results.

If you have any trouble using our code or find a bug or a mistake in this manual, please file a Github issue.