Advancing Ship Re-Identification in the Wild: The ShipReID-2400 Benchmark Dataset and D2InterNet Baseline Method

Abstract

Ship Re-Identification (ReID) aims to accurately identify ships with the same identity across different times and camera views, playing a crucial role in intelligent waterway transportation. Compared to the widely researched pedestrian and vehicle ReID, Ship ReID has received much less attention, primarily due to the scarcity of largescale, high-quality ship ReID datasets available for public access. Also, capturing an entire ship can be challenging due to their large size. The visible area of ships is also dynamic, often changing significantly because of variations in cargo loading or water waves. The unique challenges faced by ships make it difficult to achieve ideal results by directly applying existing ReID methods. To address these challenges, in this paper, we introduce ShipReID-2400, a dataset for ship ReID compiled from a real-world intelligent waterway traffic monitoring system. It comprises 17,241 images of 2,400 distinct ship identities collected over 53 months, ensuring diversity and representativeness. Then, we propose a feature Disentangle-to-Interact Network (D2InterNet) for ship ReID, designed to extract discriminative local features despite significant scale variations. Given that distinctive local details are often tiny compared to the larger, less distinctive areas of ships, D2InterNet first employs a dual-branch architecture to separately focus on global and local features. Moreover, we introduce a Part-aware Feature Perception Module (PFPM) to enhance the extraction of local features, along with a Feature Interaction Module (FIM) and a Collaborative Interaction Learning Module (CILM) to effectively interact and integrate global and local features across different scales. Extensive experimental results show that D2InterNet achieves state-of-the-art ship ReID performance on both the ShipReID-2400 and VesselReID datasets. In addition, despite being designed for ship ReID, D2InterNet demonstrates comparable performance with the best methods on the MSMT17 pedestrian ReID dataset, showcasing its good generalization capability.

Examples of ships from our ShipReID-2400 (images in each column are captures of the same ship identity from different cameras). As can be seen, we not only ensure that the captured images are of high quality but also obtain images of the same ship at different locations, angles, times, and weather conditions. For ease of presentation, the images here are scaled.

Pipeline

The overall architecture of our proposed D2InterNet. To address significant scale variation and extract discriminative local features for ship ReID, D2InterNet first uses a dual-branch architecture to independently focus on global and local features, enhanced by a Part-aware Feature Perception Module (PFPM) for better local feature capture. We then introduce Feature Interaction Modules (FIM) and Collaborative Interaction Learning Module (CILM) to interact and integrate features across scales. D2InterNet enables flexible dual-branch training and single-branch testing without additional computational overhead.

Installation

Our code operates in an environment with Python 3.8 and Torch 1.9.1, you can use the following code for installation. Please ensure to select the appropriate CUDA version.

conda create -n shipreid python=3.8
conda activate shipreid
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Dataset

Before formally accessing the dataset, applicants must sign a dataset license agreement(中国境内的申请者请签署和发送中文版数据集版权协议).

Please print the license agreement, hand-sign it, scan the signed document, and email it to us from your official institutional email address to apply for dataset access (mailto: liubaolongx@gmail.com or dongjf24@gmail.com).

Since our data is stored on Google Drive, you will first need to register for a Google Drive account.

Then, please send the dataset license agreement along with your Google Drive account details to us via the application email. Upon approval of the license agreement, we will grant your account access to the dataset.

Image Naming Convention

shipId_camId_idx_imgName

e.g.: 2106_c003_01_T_20180703_15_38_50_734375.jpg

shipId	camId	Idx	imgName
2106	c003	01	T_20180703_15_38_50_734375.jpg

shipId: Assigned based on ship license plate. Different plates correspond to different shipIds. Range: 0 <= shipId < 2400
camId: Assigned based on camera location. Different locations correspond to different camIds. Range: 1 <= camId <= 8
Idx: Sequence number for images of the same shipId at the same camId. Example: The first image of shipId 2106 at location c003 is "01"
imgName: Original image filename containing the image capture time

Dataset Partition

The 2400 ship IDs are randomly shuffled and split into train, validation, and test sets in a 0.75:0.125:0.125 ratio:

• Randomly select 1800 IDs → Training Set
• Randomly select 300 IDs → Validation Set
• Randomly select 300 IDs → Test Set

Sets are mutually exclusive. During training:

• Only the training set is used for weight updates
• Validation and test sets are solely for accuracy evaluation (no impact on training)

According to the re-identification task definition, both validation and test sets require corresponding query sets:

• Query images are selected as one image per ID per camera to simulate real-world query scenarios
• Validation query set: 691 images (300 IDs)
• Test query set: 706 images (300 IDs)

Set	# ID	# Images
bounding_box_train (Train)	1800	12,988
val_query (Validation Query)	300	691
test_query (Test Query)	300	706
bounding_box_test (Gallery)	600	4,253

Code Structure

Our code structure is as follows.

ShipReID-2400
├── config     # Initial configuration file definitions
├── configs    # Configuration file directory, storing YML configuration files
├── datasets   # Dataset loading and processing module
├── exp        # Experiment-related files
├── fig        # Figures
├── loss       # Loss function definitions
├── model      # Model definitions and implementations
├── processor  # Training and testing code processors
├── solver     # Optimizers and learning rate schedulers
├── utils      # Utility functions and tools
├── test.py    # Test code
└── train.py   # Train code

Training

You can use the following code for training. Please note that PART_H and PART_W represent the height and width of the local region, and PART_RATIO denotes the percentage retained by the FIM module. Ensure that DATASETS.ROOT_DIR is replaced with the path to your dataset.

# bash exp/train.sh

CUDA_VISIBLE_DEVICES=2 python ../train.py --config_file ../configs/ship/vit_base.yml \
SOLVER.EVAL_PERIOD 20 SOLVER.IMS_PER_BATCH 48 SOLVER.MAX_EPOCHS 200 \
OUTPUT_DIR ./logs/ DATASETS.NAMES "('ShipReID2400')" \
MODEL.PART_H 16 MODEL.PART_W 4 MODEL.ID_LOSS_WEIGHT 0.5 MODEL.TRIPLET_LOSS_WEIGHT 5.0 \
MODEL.PART_ID_LOSS_WEIGHT 0.5 MODEL.PART_TRIPLET_LOSS_WEIGHT 5.0 MODEL.TOKEN_CONTRAST_LOSS_WEIGHT 10.0 \
MODEL.PART_RATIO 0.6 MODEL.TOKEN_CONTRAST_TYPE 'triplet' SOLVER.TOKEN_MARGIN 0.3

Evaluation

You can use the following code for evaluation. Please note that PART_H and PART_W should correspond to those used during training, and TEST.WEIGHT is the path where the weights are stored.

# bash exp/test.sh

CUDA_VISIBLE_DEVICES=2 python ../test.py --config_file ../configs/ship/vit_base.yml \
OUTPUT_DIR ./logs/ DATASETS.NAMES "('ShipReID2400')" \
MODEL.PART_H 16 MODEL.PART_W 4 \
TEST.WEIGHT 'path_to_test_weight'