[Paper Review] Local Map Descriptor for Compressive Change Retrieval

Tanaka Kanji|arXiv (Cornell University)|Mar 3, 2016

Genomics and Phylogenetic Studies3 citations

TL;DR

This paper proposes a Local Map Descriptor (LMD) with a compressed bag-of-words (BoW) structure for scalable compressive change retrieval in robotic mapping. By using a viewpoint planner to stabilize local map origins via scene parsing, the method enables direct comparison of both appearance and spatial layout of visual features, significantly improving self-localization and change detection accuracy over traditional BoW methods, especially in non-stationary environments with loop-closing challenges.

ABSTRACT

Change detection, i.e., anomaly detection from local maps built by a mobile robot at multiple different times, is a challenging problem to solve in practice. Most previous work either cannot be applied to scenarios where the size of the map collection is large, or simply assumed that the robot self-location is globally known. In this paper, we tackle the problem of simultaneous self-localization and change detection, by reformulating the problem as a map retrieval problem, and propose a local map descriptor with a compressed bag-of-words (BoW) structure as a scalable solution. We make the following contributions. (1) To enable a direct comparison of the spatial layout of visual features between different local maps, the origin of the local map coordinate (termed "viewpoint") is planned by scene parsing and determined by our "viewpoint planner" to be invariant against small variations in self-location and changes, aiming at providing similar viewpoints for similar scenes (i.e., the relevant map pair). (2) We extend the BoW model to enable the use of not only the appearance (e.g., polestar) but also the spatial layout (e.g., spatial pyramid) of visual features with respect to the planned viewpoint. The key observation is that the planned viewpoint (i.e., the origin of local map coordinate) acts as a pseudo viewpoint that is usually required by spatial BoW (e.g., SPM) and also by anomaly detection (e.g., NN-d, LOF). (3) Experimental results on a challenging "loop-closing" scenario show that the proposed method outperforms previous BoW methods in self-localization, and furthermore, that the use of both appearance and pose information in change detection produces better results than the use of either information alone.

Motivation & Objective

• To address the challenge of scalable change detection in large-scale map collections where robot self-location is not globally known.
• To enable robust self-localization and change detection by reformulating the problem as a map retrieval task.
• To develop a local map descriptor that captures both appearance and spatial layout of features relative to a stable, scene-parsed viewpoint.
• To reduce false positives in change detection by minimizing sensitivity to viewpoint variations and sensor noise.
• To demonstrate the superiority of combining appearance and pose information in change detection over using either alone.

Proposed method

• A viewpoint planner uses scene parsing to determine a stable origin (viewpoint) for each local map, invariant to small self-location shifts and environmental changes.
• The proposed Local Map Descriptor (LMD) extends the bag-of-words model to incorporate spatial layout using a spatial pyramid structure relative to the planned viewpoint.
• The method uses a compressed BoW representation to enable efficient retrieval across large map collections.
• Anomaly detection is performed using NN-d (Nearest Neighbor distance) with a thresholded anomaly-ness score to distinguish real changes from noise or viewpoint variations.
• The approach is evaluated on 2D pointset maps from laser data, with scene parsing based on a Manhattan world grammar.
• Self-localization is achieved by retrieving the most similar database map using map-level similarity, followed by change detection on the top-ranked map.

Experimental results

Research questions

• RQ1 Can a viewpoint planner based on scene parsing stabilize local map origins to enable direct comparison of visual features across different robot poses?
• RQ2 Does incorporating spatial layout information into the bag-of-words model improve change detection performance compared to appearance-only BoW?
• RQ3 How does the proposed LMD method perform in self-localization and change detection under challenging loop-closing scenarios with unknown self-location?
• RQ4 Is the combination of appearance and pose information more effective than using either alone for detecting meaningful environmental changes?
• RQ5 How does the method scale in performance when applied to large collections of local maps with dynamic environments?

Key findings

• The proposed LMD method outperformed standard BoW methods in self-localization, demonstrating improved robustness to viewpoint variations and environmental changes.
• Using both appearance and spatial layout information in the LMD descriptor led to significantly better change detection results than using either feature type alone.
• The optimal threshold for NN-d anomaly detection was found to be in the range of 1.0 to 3.0, balancing false positives and missed detections.
• Performance degraded in non-stationary environments compared to stationary ones, highlighting the difficulty of change detection when new objects appear.
• The viewpoint planner successfully generated consistent viewpoints for similar scenes, enabling reliable map comparison without pre-alignment.
• The method achieved an average ANR (Average Number of Retrievals) of 25.1 across eight descriptors, with descriptor #1 (polestar) selected for simplicity due to minimal performance differences.