Abstract
Place recognition is a fundamental capability for autonomous robots, enabling loop closure and global localization within SLAM (Simultaneous Localization and Mapping) frameworks. While deep learning models have achieved high accuracy in LiDAR-based place recognition, they typically suffer from catastrophic forgetting when deployed sequentially across different environments. This paper introduces InCloud, the first incremental learning framework specifically designed for point cloud place recognition. By employing a novel structure-aware angular distillation loss and a sigmoid-based relaxation mechanism, InCloud maintains the geometric integrity of previously learned environments while adapting to new domains. Experimental results across multiple large-scale datasets demonstrate that InCloud significantly outperforms standard incremental learning baselines and improves zero-shot generalization.
1. Introduction
For a robot to navigate effectively over long periods, it must solve the "Where am I?" problem. Deep learning models for place recognition are usually trained on static datasets; however, real-world robotics involves sequential deployment across diverse cities and sensor configurations. When a model trained on one city (e.g., Oxford) is fine-tuned on another (e.g., Singapore), it often loses its ability to recognize the original environment—a phenomenon known as catastrophic forgetting.
The challenge lies in the stability-plasticity dilemma: the model must be stable enough to retain old knowledge but plastic enough to learn new features. InCloud addresses this by treating place recognition as an incremental metric learning problem, ensuring that the embedding space evolves without collapsing previous topological relationships.
2. The InCloud Framework
InCloud moves beyond simple fine-tuning by introducing three core technical components designed for the unique requirements of point cloud embeddings.
2.1 Structure-Aware (Angular) Distillation
Traditional incremental learning often distills class probabilities (logits). However, place recognition relies on metric learning, where the relative distance between embeddings is the primary signal. Standard Euclidean distance distillation can conflict with the triplet loss used during training; if a new task requires points to be pushed further apart, Euclidean constraints penalize this necessary movement.
InCloud introduces Angular Distillation, which preserves the angles between triplets of embeddings rather than their absolute positions. This approach captures the higher-order geometric structure of the embedding space and is invariant to rotation or scaling, allowing the model to adjust the "size" of its clusters while maintaining their relative orientation.
2.2 Distillation Relaxation
To further manage the stability-plasticity trade-off, InCloud utilizes a sigmoid-based weighting scheme for the distillation loss. At the beginning of a new training task, the distillation constraint is high to protect existing knowledge. As the model encounters more data from the new environment, the constraint gradually relaxes, providing the network with the flexibility needed to capture domain-specific nuances.
2.3 Exemplar Memory
To prevent the embedding space from drifting too far, InCloud maintains a small, fixed-size buffer (K=256) of representative point cloud pairs from previous environments. These exemplars are replayed during new training phases to anchor the embedding space and maintain consistency across tasks.
3. Experimental Evaluation
3.1 Datasets and Protocols
The framework was evaluated using four major LiDAR datasets:
- Oxford RobotCar: Urban data captured via SICK 2D scanners.
- MulRan: Data from South Korea using Ouster OS1-64 sensors.
- In-house: A Singapore-based dataset using Velodyne HDL-64E.
- KITTI: Reserved for zero-shot evaluation to test generalization to unseen sensors.
The authors tested two scenarios: a 2-Step protocol (batching environments) and a 4-Step protocol (sequential arrival of four distinct datasets).
3.2 Performance Metrics
InCloud was tested across three different backbones: PointNetVLAD, LoGG3D-Net, and MinkLoc3D. In the 4-Step protocol using MinkLoc3D, InCloud reduced the forgetting metric from 22.4% (naive fine-tuning) to just 5.7%. It achieved a mean Recall@1 (mR@1) of 83.6% across all seen domains, outperforming established incremental methods like Learning without Forgetting (LwF) and Elastic Weight Consolidation (EWC).
4. Key Insights
- Relative Geometry is Robust: The success of angular distillation suggests that in place recognition, the absolute coordinates in embedding space are less critical than the topological relationships between neighbors.
- Architecture Agnosticism: Because InCloud acts as a wrapper around the loss function and training loop, it is compatible with various neural architectures, making it a versatile tool for the robotics community.
- Enhanced Generalization: Interestingly, preventing forgetting also improved zero-shot performance on the KITTI dataset by up to 11.1%. This suggests that incremental learning encourages the model to develop a more robust, universal representation of geometric features.
5. Conclusion and Future Work
InCloud provides a robust foundation for lifelong learning in robotic perception. While it successfully mitigates catastrophic forgetting, future research may explore "data-free" incremental learning to eliminate the need for an exemplar buffer in privacy-sensitive applications. Additionally, scaling the framework to handle dozens of sequential environments remains an important challenge for long-term autonomous deployment.
Summary updated