Supervised Hashing Using Graph Cuts and Boosted Decision Trees

Embedding image features into a binary Hamming space can improve both the speed and accuracy of large-scale query-by-example image retrieval systems. Supervised hashing aims to map the original features to compact binary codes in a manner which preserves the label-based similarities of the original data. Most existing approaches apply a single form of hash function, and an optimization process which is typically deeply coupled to this specific form. This tight coupling restricts the flexibility of those methods, and can result in complex optimization problems that are difficult to solve. In this work we proffer a flexible yet simple framework that is able to accommodate different types of loss functions and hash functions. The proposed framework allows a number of existing approaches to hashing to be placed in context, and simplifies the development of new problem-specific hashing methods. Our framework decomposes the into two steps: binary code (hash bits) learning, and hash function learning. The first step can typically be formulated as a binary quadratic problem, and the second step can be accomplished by training standard binary classifiers. For solving large-scale binary code inference, we show how to ensure that the binary quadratic problems are submodular such that an efficient graph cut approach can be used. To achieve efficiency as well as efficacy on large-scale high-dimensional data, we propose to use boosted decision trees as the hash functions, which are nonlinear, highly descriptive, and very fast to train and evaluate. Experiments demonstrate that our proposed method significantly outperforms most state-of-the-art methods, especially on high-dimensional data.

💡 Research Summary

The paper addresses the problem of supervised hashing for large‑scale image retrieval, where the goal is to map high‑dimensional image descriptors into compact binary codes that preserve label‑based similarity. Existing supervised hashing methods typically tie a specific form of hash function (e.g., linear perceptron, kernel, eigenfunction) tightly to a particular loss function, resulting in highly non‑convex optimization problems that are difficult to solve and hard to extend to new hash functions or loss formulations.

To overcome these limitations, the authors propose a flexible two‑step framework called FastHash. The learning process is decomposed into (1) binary code inference and (2) hash‑function learning. In the first step, the pairwise loss defined on Hamming distance or Hamming affinity (e.g., KSH, BRE, MLH losses) is shown to be equivalent to a binary quadratic problem. By carefully reformulating the loss, the resulting quadratic energy can be made sub‑modular, which enables the use of an efficient graph‑cut based block search algorithm. The algorithm partitions the training set into blocks, solves each block with a min‑cut/max‑flow routine, and iteratively refines the binary codes. This approach yields high‑quality binary codes while scaling to hundreds of thousands or millions of training points.

In the second step, the inferred binary codes are treated as target labels for a set of independent binary classification problems—one for each hash bit. Because the classification problem is decoupled from the original loss, any binary classifier can be employed. The authors focus on boosted decision trees as hash functions. Decision trees require only simple comparison operations, making them extremely fast at test time, and they handle quantized high‑dimensional inputs with little memory overhead. By using AdaBoost (or a similar boosting scheme) they construct an ensemble of shallow trees for each bit, achieving a non‑linear mapping comparable to kernel hash functions but with far lower computational cost.

The overall algorithm proceeds bit‑by‑bit: for each bit r, (i) infer the optimal binary values for that bit across all training samples using the graph‑cut based solver, (ii) train a boosted‑tree classifier to predict those values from the original features, (iii) apply the learned classifier to update the binary codes, and then move to the next bit. This “bit‑wise” optimization allows errors made in earlier bits to be compensated in later bits, improving overall retrieval performance.

Extensive experiments are conducted on several benchmark datasets (CIFAR‑10, SUN397, ImageNet) with code lengths ranging from 64 to 128 bits. The authors evaluate mean average precision (mAP), precision‑recall curves, training time, and query speed. Results show that FastHash consistently outperforms state‑of‑the‑art supervised hashing methods such as KSH, BRE, and MLH, especially when the original features are high‑dimensional (e.g., 10,000‑dimensional bag‑of‑visual‑words). In many cases, mAP improvements of 10–20 % are reported, while training time is reduced by an order of magnitude or more compared to kernel‑based approaches. At test time, the boosted‑tree hash functions are 5–10× faster than linear perceptron hash functions and dramatically faster than kernel hash functions, making the method suitable for real‑time large‑scale retrieval.

The paper’s contributions can be summarized as follows:

1. A general two‑step hashing framework that separates loss design from hash‑function design, enabling arbitrary loss functions (any Hamming‑based loss) and arbitrary hash families.
2. A sub‑modular formulation of binary quadratic losses and an efficient graph‑cut based block search algorithm for large‑scale binary code inference.
3. The introduction of boosted decision trees as fast, memory‑efficient, non‑linear hash functions that scale to high‑dimensional data.
4. Comprehensive empirical validation demonstrating superior accuracy and speed, together with publicly released code for reproducibility.

Overall, FastHash provides a practical, extensible solution for supervised hashing, opening the door to further extensions such as deep neural network classifiers for hash functions, multi‑modal similarity learning, and applications beyond image retrieval.