[Paper Review] MILD: Multi-Index hashing for Loop closure Detection

Lei Han, Lu Fang|arXiv (Cornell University)|Feb 28, 2017

Algorithms and Data Compression4 citations

TL;DR

MILD proposes a real-time loop closure detection system using Multi-Index Hashing (MIH) to enable efficient, high-recall image similarity matching via direct feature matching, avoiding the need for bag-of-words representation. It achieves state-of-the-art accuracy (up to 94.5% recall) with low latency (35ms per image), outperforming both SIFT/SURF and binary feature-based methods in speed and recall.

ABSTRACT

Loop Closure Detection (LCD) has been proved to be extremely useful in global consistent visual Simultaneously Localization and Mapping (SLAM) and appearance-based robot relocalization. Methods exploiting binary features in bag of words representation have recently gained a lot of popularity for their efficiency, but suffer from low recall due to the inherent drawback that high dimensional binary feature descriptors lack well-defined centroids. In this paper, we propose a realtime LCD approach called MILD (Multi-Index Hashing for Loop closure Detection), in which image similarity is measured by feature matching directly to achieve high recall without introducing extra computational complexity with the aid of Multi-Index Hashing (MIH). A theoretical analysis of the approximate image similarity measurement using MIH is presented, which reveals the trade-off between efficiency and accuracy from a probabilistic perspective. Extensive comparisons with state-of-the-art LCD methods demonstrate the superiority of MILD in both efficiency and accuracy.

Motivation & Objective

• Address the low recall issue in binary feature-based loop closure detection due to poor clustering in bag-of-words representations.
• Overcome the high computational cost of SIFT/SURF-based methods while maintaining high accuracy.
• Enable real-time loop closure detection with multiple loop closures per query image, unlike most prior methods that assume single loop closure.
• Develop an efficient approximate similarity measurement using MIH that balances accuracy and computational complexity.
• Provide a theoretical probabilistic analysis of MIH-based similarity approximation to guide parameter selection and ensure robust performance.

Proposed method

• Replace traditional bag-of-words representation with direct feature matching using binary ORB features.
• Apply Multi-Index Hashing (MIH) to index binary feature descriptors, enabling fast approximate nearest neighbor search.
• Use multiple hash tables with substrings of the binary descriptor (16 substrings of 16 bits each) to improve retrieval efficiency.
• Employ Bayesian inference on MIH-based similarity scores to select loop closures, incorporating confidence thresholds.
• Set a Hamming distance threshold (d₀ = 60) and a loop closure probability threshold (P₀ = 0.7) to filter candidates.
• Discard hash entries with high collision frequency to reduce memory and computation overhead without significant accuracy loss.

Experimental results

Research questions

• RQ1 Can Multi-Index Hashing enable high-accuracy, real-time loop closure detection without relying on bag-of-words representation?
• RQ2 What is the trade-off between accuracy and efficiency in MIH-based image similarity approximation, and how can it be theoretically analyzed?
• RQ3 Can MILD detect multiple loop closures per query image, and how does this compare to single-loop assumptions in prior work?
• RQ4 How does MILD’s performance in recall and latency compare to state-of-the-art SIFT/SURF and binary feature-based methods?
• RQ5 What is the memory and computational footprint of MILD, and is it feasible for deployment on mobile or embedded devices?

Key findings

• MILD achieves 87.3% recall on the NewCollege dataset and 94.5% on Lip6Indoor, outperforming BOBW (30.6% and 41.9%) and IBuILD (38% and 25.5%) in accuracy.
• MILD processes each image in 35ms on average, significantly faster than RTABMAP (700ms) and Angeli (753ms), while matching or exceeding their recall.
• The method achieves high accuracy comparable to SIFT/SURF-based methods (e.g., 98% recall in RTABMAP) but with 20x lower latency.
• Memory usage is minimal—84MB for the NewCollege dataset—making it suitable for mobile and embedded deployment.
• The theoretical analysis of MIH-based similarity shows a clear trade-off between accuracy and complexity, validated by empirical results.
• MILD successfully detects multiple loop closures per query, demonstrating robustness in complex, repeated environment scenarios.