Triplet Loss for Person Re-Identification

• TensorFlow implementation of person re-identification using triplet loss described in In Defense of the Triplet Loss for Person Re-Identification .
• In this repository, the metric learning using triplet loss is tested on MNIST and Market-1501 Dataset. For both dataset, batch hard mining is used during training, which is implemented in src/triplet/hard_mining.py.
• Below are some results on Market-1501 Dataset. From left to right: query image and top 10 retrieved images (green boarder - true positive; red border - false positive).

Requirements

• Python 3.0
• TensorFlow 1.12.0+
• sklearn

Metric Learning on MNIST

• The aim of this experiment is to check if the triplet loss using batch hard mining is correctly implemented.
• The dimension of embedding is set to be 2 for visualization.
• As MNIST dataset is pretty simple, the network used to learn the embedding consists of 2 convolutional layers followed by 2 fully connect layers.
• For each batch, 12 samples of each class (0-9) are randomly selected. For triplet loss, margin = 0.5 is used.
• Below is the learned embedding of testing set after training for 50 epochs with learning rate 1e-4. We can see different classes are well grouped.

Train embedding of MNIST by yourself

Setup path

• All the paths are setup in experiment/config.py.
• mnist_dir - directory of MNIST dataset.
• mnist_save_path - directory of saving results.

Train

Go to experiment, run

python train.py --dateset mnist --lr 1e-4 \
  --embed EMBEDDING_DIM \
  --margin MARGIN_VAL

• --embed is for dimension of embedding and --margin is for margin value of triplet loss
• If dimension of embedding is 2, the embedding visualization images will be saved in mnist_save_path for each epoch.

Person Re-Identification on Market-1501 Dataset

• The network for learning embedding is the same as LuNet described in In Defense of the Triplet Loss for Person Re-Identification .
• Images are rescaled to 128 x 64 and the intensity is normalized to [0, 1] before fed into the network.
• Batch size is 128 (randomly pick 32 identities with 4 images each). Embedding size is 128.
• Below is the training-log for 25000 training iteration. The initial learning rate is 5e-4 and margin = 0.5 is used. Row 1: Left: non-zero loss ratio of hardest triplet at each training step; Right: averaged hardest triplet loss over 100 training steps at every 100 steps. Row 2: Left: statistics of embedding entries; Right: statistics of 2-norm of embeddings.

• Re-ranking is used when retrieving images from gallery set given a set of query images.
• Below are retrieve results given query images. From left to right: query image and top 10 retrieved images (green boarder - true positive; red border - false positive).

Train and inference embedding for person re-identification on Market-1501 Dataset

Preparation

• Download the dataset from here.
• Setup in experiment/config.py:
• market_dir - directory of dataset.
• market_save_path - directory of saving results.

Train

Go to experiment, run

python train.py --dataset market --lr 5e-4 \
  --embed EMBEDDING_DIM \
  --margin MARGIN_VAL \
  --folder FOLDER_NAME

• --embed is for dimension of embedding, --margin is for margin value of triplet loss and --folder is for subfolder under market_save_path used for separating results of different experiments.
• Summary file for Tensorboard visualization will be saved in market_save_path/FOLDER_NAME/.
• The trained model will be saved in market_save_path/FOLDER_NAME/ every 5000 training steps.

Inference and visualize ranking

Go to experiment, run

python viz_ranking --dataset market \
  --embed EMBEDDING_DIM \
  --top_k TOP_K \
  --n_query NUM_QUERY \
  --folder FOLDER_NAME \
  --load_id MODEL_ID 

• --load_id is for the index (training step when saved) of trained model to be loaded, --n_query is for the number of query images used for visualization and --top_k is for how many closest images in gallery set are selected as retrieved images for each query image.
• NUM_QUERY query images will be randomly selected from testing set and rest of images are regarded as gallery set. Embedding of entire testing set (19732 images) will be computed before ranking.
• Result images will be saved in market_save_path/FOLDER_NAME/.
• The mAP increased around 10% after re-ranking is used. The re-rankning and evaluation code are borrow from here currently.

Reference code

• https://github.com/omoindrot/tensorflow-triplet-loss
• https://github.com/huanghoujing/person-reid-triplet-loss-baseline

Author

Qian Ge