[Paper Review] The 2021 Hotel-ID to Combat Human Trafficking Competition Dataset

Rashmi Kamath, Gregory Rolwes+2 more|arXiv (Cornell University)|Jun 10, 2021

Sex work and related issues12 references5 citations

TL;DR

This paper introduces the 2021 Hotel-ID dataset, a large-scale, real-world benchmark for fine-grained visual recognition in the context of human trafficking investigations. It presents a challenge to identify hotels from low-quality, occluded, and diverse-angle images collected via the TraffickCam mobile app, with a classification-based approach achieving SOTA performance (R@1: 49.3%) while metric learning methods show better generalization potential for open-world deployment.

ABSTRACT

Hotel recognition is an important task for human trafficking investigations since victims are often photographed in hotel rooms. Identifying these hotels is vital to trafficking investigations since they can help track down current and future victims who might be taken to the same places. Hotel recognition is a challenging fine grained visual classification task as there can be little similarity between different rooms within the same hotel, and high similarity between rooms from different hotels (especially if they are from the same chain). Hotel recognition to combat human trafficking poses additional challenges as investigative images are often low quality, contain uncommon camera angles and are highly occluded. Here, we present the 2021 Hotel-ID dataset to help raise awareness for this problem and generate novel approaches. The dataset consists of hotel room images that have been crowd-sourced and uploaded through the TraffickCam mobile application. The quality of these images is similar to investigative images and hence models trained on these images have good chances of accurately narrowing down on the correct hotel.

Motivation & Objective

• To address the critical need for accurate hotel recognition in human trafficking investigations, where victims are often photographed in hotel rooms.
• To provide a realistic, large-scale benchmark for fine-grained visual recognition under real-world constraints such as low image quality, occlusion, and variable camera angles.
• To raise awareness and stimulate novel research in visual recognition for social good, particularly in combating human trafficking.
• To develop a dataset and competition that supports both high-accuracy classification and scalable, generalizable retrieval methods for new hotels.

Proposed method

• The dataset is constructed from crowd-sourced images uploaded via the TraffickCam mobile application, capturing real hotel rooms from travelers worldwide.
• The training set contains 97,527 images from 7,770 unique hotels across 86 hotel chains, with images reflecting real-world variability in lighting, composition, and occlusion.
• The test set includes 12,400 images from hotels present in the training set, captured by different users and devices to simulate real-world deployment.
• Baseline models are evaluated using both classification (cross-entropy loss) and metric learning (SCT, EPHN, Batch-All) approaches to compare performance and generalization.
• Evaluation uses standard retrieval metrics: Recall@K (K=1, 10, 100) and Mean Average Precision at 5 (MAP@5), with results reported on the test set.
• The system is designed with ethical safeguards, restricting access to the National Center for Missing and Exploited Children and minimizing data exposure to reduce misuse risks.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on real-world, low-quality hotel images achieve high accuracy in identifying the correct hotel from a test image?
• RQ2 How do classification-based models compare to metric learning-based models in terms of performance and generalization to unseen hotels?
• RQ3 To what extent do visual similarities between rooms from the same hotel chain hinder or help recognition, especially when rooms from different hotels in the same chain are visually more similar?
• RQ4 Can models trained on this dataset generalize to new hotels not seen during training, particularly in a real-world deployment setting with hundreds of thousands of potential hotel classes?
• RQ5 What are the ethical implications of deploying hotel recognition systems, and how can they be mitigated to prevent misuse against vulnerable populations?

Key findings

• The cross-entropy classification-based approach achieved the highest performance, with a Recall@1 score of 49.3%, significantly outperforming all metric learning baselines.
• Metric learning methods such as SCT (Selectively Contrastive Training) achieved a Recall@1 of 36.04%, showing strong potential for generalization to new hotels in open-world settings.
• The classification model achieved a Mean Average Precision at 5 (MAP@5) of 55.1%, indicating strong top-5 retrieval performance despite the fine-grained nature of the task.
• Even in challenging cases with high occlusion or unusual camera angles, the classification model found relevant matches, including nearly identical rooms from different hotels in the same chain.
• Qualitative analysis showed that metric learning methods like SCT were more sensitive to specific visual cues such as carpets or shelving, while the classification model was more robust overall.
• The study highlights the trade-off between high accuracy on known hotels and scalability to new hotels, with metric learning methods being better suited for real-world deployment.