Database
What Should Visual Place Recognition Be Invariant To? Rethinking Condition Invariance in Visual Place Recognition
arXiv cs.RO0 views
arXiv:2608.06847v2 Announce Type: replace Abstract: Condition invariance in Visual Place Recognition (VPR) is typically evaluated by matching queries from one condition against a database from another condition. This tests the ability to recognize the same place when the query's condition differs from the database's condition. However, to map large environments, a database can be built from images collected by multiple robots or across multiple traversals under different conditions. In this setting, condition blindness---the ability to remain invariant to condition differences among database images---becomes important, but standard cross-condition evaluation does not provide a controlled assessment of this ability. Therefore, we introduce Distractor-Augmented Recall (DAR) to measure condition blindness in a controlled manner. Across eleven methods and six datasets, we find that differences in condition blindness can be larger than Recall@1 alone would suggest. To improve the condition blindness that DAR reveals, we compare three methods that make VPR descriptors more invariant to condition changes through post-processing. Among the three methods, we demonstrate INLP on a database built from multiple traversals under different conditions, showing that it can improve the performance of existing VPR methods.
Source:https://arxiv.org/abs/2608.06847PreviousNeural LiDAR Bundle AdjustmentNextProjection-Retraction MPPI: Exact Command-Space Constraint-Manifold Control for Manipulators
Related posts
TimescaleDB vs ClickHouse: Real-Time Time-Series Ingestion, Compression & Query Latency
0dev.to AI
Pinecone vs pgvector on AWS RDS: The Exact Vector Scale and QPS Inflection Point Where Managed Vector DBs Cost 4x More
0dev.to AWS

Two Specialist Agents, One Incident: What an A2A Handoff Takes
0dev.to AWS
From v6.0 to v6.5.0: everything I changed in my post-quantum messenger
0dev.to Security

One VS Code panel for my whole local stack: Docker, Kubernetes, Postgres and Kafka
0dev.to Docker
Shipped a Yamtrack one-click deploy template — self-hosting media requests without the SaaS
0dev.to Docker