Pith. sign in

REVIEW 5 major objections 5 minor 61 references

LifelongPR: Lifelong point cloud place recognition based on sample replay and prompt learning

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LifelongPR claims that one point-cloud place-recognition model can keep learning new cities and sensors without forgetting old ones, using information-aware replay and a small prompt module.

desk verdict A useful continual-learning recipe for point cloud place recognition, but the central SOTA claim is contradicted by its own tables and the abstract cherry-picks across configurations. read the letter →

arxiv 2507.10034 v2 pith:AQX5OKJM submitted 2025-07-14 cs.CV cs.RO

Xianghong Zou , Jianping Li , Zhe Chen , Zhen Cao , Zhen Dong , Qiegen Liu , Bisheng Yang This is my paper · ORCID
classification cs.CVcs.RO
keywords pointcloudplacerecognitioncontinuallearningcatastrophicforgettingreplaysampleselectionpromptLiDARdomainshiftincremental
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

LifelongPR is a framework for continual learning in point-cloud place recognition: the task of updating a recognition model sequentially on new cities, scenes, and LiDAR sensors without forgetting the places it learned earlier. The paper argues that the two standard remedies—replaying old samples and distilling old knowledge—fail because replay memory is chosen randomly and because small replay buffers cannot bridge large domain shifts. Its method allocates a fixed 256-sample replay memory across datasets according to each dataset's estimated information content, selects spatially diverse samples, and adds a lightweight prompt module trained in two stages to carry domain-specific knowledge. On two four-dataset sequences the reported numbers beat the previous state of the art, including a 6.50% gain in mean incremental top-1 recall and a 7.96% gain in mean recall on the harder sequence with the MinkLoc3D network, plus an 8.95-point drop in the forgetting score on that sequence with PointNetVLAD. If the claim holds, a single incrementally updated model could replace per-city retraining or maintaining separate per-domain models.

What carries the argument

The load-bearing machinery is a pair of mechanisms targeting the two failure modes the paper identifies. Replay selection: for each historical training set, the information quantity is estimated as the effective rank of a Gaussian kernel matrix over that set's features, normalized to [0,1]; a temperature-controlled softmax then allocates the fixed total replay budget of 256 submaps among old and new datasets, and a greedy algorithm picks samples that maximize the diversity score g(M)=sum of per-sample minimum combined distances in Euclidean and feature space. This determines both how many and which old scenes are replayed. Prompt module: a lightweight attention module with learnable prompts as queries (following the Q-Former design) computes domain-specific auxiliary features from the raw point cloud, which are fed into the backbone without modifying it; a two-stage training schedule—stage 1 fits the prompts with the backbone frozen, stage 2 fine-tunes the backbone with the prompts frozen—lets the model adapt to a new domain while retaining old knowledge.

What would settle it

Run the full LifelongPR pipeline on Seq2 with the MinkLoc3D network but replace the information-aware replay selection with 256 randomly chosen past submaps, keeping the prompt module and two-stage training; if mIR@1 and mR@1 remain at the reported 79.67% and 75.08%, the replay-selection component is not doing the claimed work.

Watch

Extended reading notes

Core claim

The central claim is that catastrophic forgetting in point-cloud place recognition can be substantially reduced by pairing two complementary mechanisms rather than by more aggressive regularization or contrastive training. LifelongPR keeps a fixed-size memory of past submaps, but unlike random replay it decides how many slots each old dataset gets by estimating the effective rank of a Gaussian kernel matrix over that dataset's features, and it chooses which scenes to store with a greedy maximization of a spatial-diversity score that penalizes closeness in both Euclidean and feature space. To handle domain shift between tasks, a lightweight Q-Former-style prompt module with learnable prompts acting as attention queries injects domain-specific auxiliary features into the backbone, and a two-stage training schedule first fits the prompts with the backbone frozen, then fine-tunes the backbone with the prompts frozen. Across two dataset sequences—Oxford to DCC to Riverside to In-house, and the harder Oxford to Hankou to Campus to In-house—and across PointNetVLAD, PatchAugNet, and MinkLoc3D backbones, the authors report that LifelongPR consistently matches or outperforms the InCloud and CCL baselines; on Seq2 with MinkLoc3D the reported values are 79.67% mIR@1, 75.08% mR@1, and an 11.42 forgetting score, against 73.16%, 67.12%, and 12.01 for CCL.

Load-bearing premise

The load-bearing premise is that the feature-extraction network already has strong, general-purpose place-recognition abilities before continual learning starts; the paper itself notes that prompt-based continual learning requires the network to be initially trained for robust feature extraction and generalizability.

Editorial extensions

If this is right

  • A single place-recognition model can be updated sequentially across new cities and LiDAR types, replacing per-city retraining or separate per-domain models.
  • On the harder Seq2 benchmark, LifelongPR with MinkLoc3D reaches 75.08% mean top-1 recall on all seen datasets after the last stage, versus 67.12% for CCL.
  • After continual learning on Seq1, Oxford top-1 recall falls only 7.20% from its initial level with LifelongPR, compared with 23.10% for fine-tuning and 10.90% for CCL.
  • Replay memory of 256 submaps is used more efficiently: datasets with higher estimated information content receive more replay slots, and selected scenes are dispersed in both Euclidean and feature space.
  • The prompt module works without modifying the backbone, so the same continual-learning wrapper can be attached to PointNetVLAD, PatchAugNet, and MinkLoc3D alike.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper does not explore: the replay-selection rule is stated over arbitrary features, so it could be lifted into other replay-based continual learners such as image retrieval or metric learning.
  • A direct stress test suggested by the paper's own limitation: vary the initial network's pretraining quality (less data, fewer epochs, random initialization) and measure how quickly the reported mIR@1 and mR@1 gains shrink.
  • Since the paper reports that the prompt module itself suffers catastrophic forgetting, a natural follow-up is to apply the same information-aware replay selection to old prompts, not only to old point clouds.
  • The fixed 256-sample replay budget is never swept; testing larger and smaller budgets would show whether the margin over CCL depends on memory pressure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes LifelongPR, a continual learning framework for point cloud place recognition (PCPR). It combines two ideas: a replay sample selection method that allocates replay budget by an information-quantity estimate and selects spatially/feature-diverse samples via a greedy algorithm, and a prompt-learning module with a two-stage training strategy that adapts non-transformer PCPR backbones to new domains while mitigating catastrophic forgetting. The method is evaluated on two four-dataset sequences (Seq1, Seq2) with three backbones (PointNetVLAD, PatchAugNet, Minkloc3D) against FT, InCloud, and CCL, with additional ablations, parameter analyses, visualizations, and case studies.

Significance. If the reported results hold, the work is a useful contribution to continual learning for PCPR: it directly targets two known limitations of existing PCPR-specific CL methods (uninformative random replay and poor handling of domain shifts), and it extends prompt-based CL to sparse-convolution and PointNet-style architectures rather than transformer backbones. Public release of code and pretrained models as well as a self-collected heterogeneous dataset are concrete strengths. However, the paper's headline claim of consistent state-of-the-art performance is contradicted by its own tables, and the lack of uncertainty quantification makes the magnitude of the claimed gains difficult to assess. The component ideas are plausible and worth further evaluation after the claims and experiments are tightened.

major comments (5)
  1. [Section IV-B, Tables II and III] The claim that LifelongPR "consistently achieves the best CL performance across different backbone networks and datasets of varying difficulty levels" is not supported by the paper's own numbers. In Table II, with PatchAugNet on Seq1, LifelongPR has mIR@1 = 85.70% versus CCL's 86.22%; in Table III, with PointNetVLAD on Seq2, LifelongPR has mIR@1 = 59.47% versus CCL's 60.18%; and in Table II, with Minkloc3D on Seq1, the F metric is worse (5.98% versus 4.40%). In addition, the three headline gains in the abstract are not realized by any single configuration: the +6.50% mIR@1 and +7.96% mR@1 come from the Minkloc3D row of Seq2, while the -8.95% F comes from the PointNetVLAD row of Seq2. The text should report per-configuration results and revise the "consistent" claim to match the evidence.
  2. [Section III-B, Eq. (5)] Equation (5) allocates replay sizes using a softmax over \sum_{i=1}^{T} exp(InfoQ(D_i)/\tau), but at the time dataset D_t is learned, the future datasets D_{t+1},...,D_T are not available in the described incremental setting. Algorithm 1 also invokes Eq. (5) after computing InfoQ only for D_1,...,D_t. Please clarify whether the allocation is computed online with T replaced by the current stage index t, or whether the full sequence of InfoQ values is assumed known in advance. If the latter, the procedure is not incremental as presented, and the replay allocation step needs to be reformulated or its assumptions stated explicitly.
  3. [Section V-E] The limitations paragraph concedes that the prompt module itself exhibits catastrophic forgetting and that the framework requires a backbone initially trained to have robust feature extraction and generalizability. Since the central claim is about lifelong retention, the current evidence on only three-task sequences does not establish that forgetting in the prompt module or the backbone remains bounded over longer sequences. At minimum, the authors should report results on longer task sequences and an ablation in which the initial backbone is trained less extensively, so readers can see how the claimed gains depend on the initialization and on the number of tasks.
  4. [Section IV, experimental setup] No standard deviations, number of runs, random seeds, or significance tests are reported for any table. Several advantages over CCL are under one percentage point (e.g., PatchAugNet on Seq2, mIR@1 +0.27%; Minkloc3D on Seq1, mR@1 +0.57%), so the "consistent best" claim may reflect training noise. Please report multiple runs with variance and, where feasible, paired significance tests, and calibrate the wording of the claims to the level of uncertainty.
  5. [Section V-D, Table VI] The temperature \tau is selected as the value giving the best mIR@1, mR@1, and F in Table VI, and those same metrics are later reported for the final method in Tables II and III. Unless \tau was chosen on a validation split, the reported gains may be optimistic due to test-set-based hyperparameter selection. Please clarify the selection procedure or re-tune \tau on a held-out split.
minor comments (5)
  1. [Section III-A, Eqs. (13)-(14)] In Eq. (14), the summand should be mR@1_t rather than mR@1_k; as written, the index k appears both as the upper limit and inside the summand, making the metric ill-defined.
  2. [Abstract and title page] The code URL differs between the abstract (github.com/zouxianghong/LifelongPR) and the full text (zouxianghong.github.io/LifelongPR); please unify the URLs.
  3. [Captions of Figs. 6 and 7] Both captions list two panels labeled "(d)"; the last panel should be labeled "(e)". Similarly, Fig. 8 lists "(d)" but omits "(c)".
  4. [Section III-B, Eqs. (3)-(4)] Please report the actual InfoQ values and a sensitivity analysis over the effective-rank threshold \epsilon and the kernel bandwidth \gamma. With \epsilon = 10^{-6}, the effective rank may often be close to n, which would make the allocation mechanism nearly uniform and should be checked against the claimed allocation behavior in Fig. 11.
  5. [Section IV-A] A joint-training upper bound (training on all datasets simultaneously) is not reported; since the paper frames joint retraining as the expensive alternative, adding this oracle would help calibrate how much of the performance gap remains.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity: the paper's contributions are validated by ablations and independent comparisons, and its self-citations are not load-bearing.

full rationale

LifelongPR is an empirical method paper. The replay allocation in Eq. (5) and the greedy selection in Eqs. (6)-(8) are proposed mechanisms, not derived predictions; their contribution is tested by ablations (Tables IV-V) against random replay and alternative training schedules. The prompt module and two-stage strategy are likewise validated empirically rather than being entailed by their definitions. The choice of tau=4.0 in Table VI is a hyperparameter selection made on the same Seq2 benchmark, which raises a statistical optimism concern but is disclosed in a parameter analysis and does not make the reported gains equal to a fitted value by construction. Self-citations, notably PatchAugNet [13] for a backbone and dataset, are not used to justify the central CL claim; the backbone is compared as a baseline without its patch augmentation module. The stated limitation in Section V-E that the backbone must already be robust is an acknowledged assumption, not a circular step. The inconsistency between the 'consistently best' claim and Tables II-III concerns correctness and statistical support, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method depends on several hand-chosen hyperparameters, an unproven rank-based information measure, and a prompt architecture borrowed from image-language models. These are not derived from first principles, and some are tuned on the test data. The replayed sample selection, the central contribution, rests on the validity of the InfoQ proxy and the greedy heuristic.

free parameters (6)
  • gamma (RBF kernel bandwidth) = 0.2
    Hand-chosen threshold in Eq. (1) controlling the Gaussian kernel used to compute InfoQ; no sensitivity analysis is reported.
  • epsilon (effective rank threshold) = 1e-6
    Hand-chosen threshold in Eq. (3) for counting singular values; affects the InfoQ estimate of each dataset.
  • tau (softmax temperature) = 4.0
    Temperature in Eq. (5) for replay size allocation; tuned on test results in Table VI without a separate validation split.
  • dthr (Euclidean distance threshold) = 1e3 m
    Hand-chosen threshold in Eq. (7) for spatial diversity selection; no sensitivity analysis.
  • alpha (candidate subset ratio) = not specified
    Ratio in Eq. (9) controlling the random subset size in the greedy algorithm; the value is not reported, affecting reproducibility.
  • ktotal (replay memory size) = 256
    Fixed total number of replay samples in Eq. (5); no experiments vary this budget.
assumptions (5)
  • ad hoc to paper The effective rank of the Gaussian kernel matrix on sample features measures the information quantity of a training set.
    Remark 1 and Sec III-B1 introduce this as the basis for replay sample allocation; no formal justification or external benchmark is provided.
  • ad hoc to paper The greedy algorithm with random candidate subsets solves the subset selection problem in Eq. (6) sufficiently well.
    Sec III-B2 replaces the NP-hard optimization with a greedy heuristic and a random subset R, without optimality guarantees.
  • ad hoc to paper A prompt module with 64 prompts and 2 attention modules can capture domain-specific knowledge when inserted into non-transformer PCPR backbones.
    Sec III-C adopts the QFormer-style architecture and fixes kq=64, d=8, N=2 based on prior work, without ablating these choices.
  • ad hoc to paper The two-stage training strategy (freeze backbone then freeze prompts) leads to stable domain-adaptive features.
    Sec III-C and V-C assert this yields better results than one-stage or prompt-only training, but the mechanism is not analyzed further.
  • domain assumption The triplet loss plus knowledge distillation with a decreasing weight schedule from InCloud is suitable for this continual learning setup.
    Sec III-D states the loss is consistent with InCloud [19]; the schedule and weights are not fully specified in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LifelongPR: Lifelong point cloud place recognition based on sample replay and prompt learning." pith.science (2026). https://pith.science/paper/AQX5OKJM

@misc{pith2026250710034,
  author       = {Pith},
  title        = {Pith review of: LifelongPR: Lifelong point cloud place recognition based on sample replay and prompt learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQX5OKJM}},
  note         = {Machine review of arXiv:2507.10034}
}
read the original abstract

Point cloud place recognition (PCPR) determines the geo-location within a prebuilt map and plays a crucial role in geoscience and robotics applications such as autonomous driving, intelligent transportation, and augmented reality. In real-world large-scale deployments of a geographic positioning system, PCPR models must continuously acquire, update, and accumulate knowledge to adapt to diverse and dynamic environments, i.e., the ability known as continual learning (CL). However, existing PCPR models often suffer from catastrophic forgetting, leading to significant performance degradation in previously learned scenes when adapting to new environments or sensor types. This results in poor model scalability, increased maintenance costs, and system deployment difficulties, undermining the practicality of PCPR. To address these issues, we propose LifelongPR, a novel continual learning framework for PCPR, which effectively extracts and fuses knowledge from sequential point cloud data. First, to alleviate the knowledge loss, we propose a replay sample selection method that dynamically allocates sample sizes according to each dataset's information quantity and selects spatially diverse samples for maximal representativeness. Second, to handle domain shifts, we design a prompt learning-based CL framework with a lightweight prompt module and a two-stage training strategy, enabling domain-specific feature adaptation while minimizing forgetting. Comprehensive experiments on large-scale public and self-collected datasets are conducted to validate the effectiveness of the proposed method. Compared with state-of-the-art (SOTA) methods, our method achieves 6.50% improvement in mIR@1, 7.96% improvement in mR@1, and an 8.95% reduction in F. The code and pre-trained models are publicly available at https://github.com/zouxianghong/LifelongPR.

Figures

Figures reproduced from arXiv: 2507.10034 by the authors.

Figure 1
Figure 1. Lifelong point cloud place recognition encounters challenges posed by diverse localization scenes and LiDAR sensors. Above information mainly [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. (a) overall workflow, (b) replay sample selection, (c) prompt module. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Replay-based continual learning for point cloud place recognition. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Location of the prompt module within the backbone network. (a) PointNetVLAD, (b) Minkloc3D, (c) PatchAugNet. Red arrows denote the insertion [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Two-stage training strategy. (a) stage 1, (b) stage 2. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Average Recall@1 on Seq1. (a) FT, (b) InCloud, (c) CCL (SOTA method), (d) LifelongPR (ours), (d) improvements of LifelongPR compared to CCL. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Average Recall@1 on Seq2. (a) FT, (b) InCloud, (c) CCL (SOTA method), (d) LifelongPR (ours), (d) improvements of LifelongPR compared to CCL. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Place recognition results on Oxford after continual learning on Seq1. (a) FT ( [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: t-SNE visualization of samples in Seq2 at different stages of [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Three success cases of LifelongPR on Seq2. Column 1: query; Column 2: top 1 submaps retrieved by the model trained on Oxford; Columns 3-4: [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Replay sample size allocation results. (a) not use information [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Average Recall@1 results of LifelongPR with different training strategies. (a) no prompt, (b) one-stage, (c) prompt only, (d) two-stage (ours). [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Three bad cases of LifelongPR trained on Seq1 and Seq2. Column 1: query; Column 2: top 1 submaps retrieved by the model trained on Oxford; [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 55 canonical work pages

  1. [1]

    J. Zhou, C. Long, Y . Xie, J. Wang, C. Zhang, B. Li, H. Wang, Z. Chen, and Z. Dong, “Whu-synthetic: A synthetic perception dataset for 3d multi-task model JOURNAL OF LATEX CLASS FILES, VOL. XXX, NO. XXX, XXX XXX 13 Fig. 12. Average Recall@1 results of LifelongPR with different training strategies. (a) no prompt, (b) one-stage, (c) prompt only, (d) two-sta...

    work page 2025

  2. [2]

    Salien- cyi2ploc: saliency-guided image-point cloud localization using contrastive learning,

    Y . Li, J. Li, Z. Dong, Y . Wang, and B. Yang, “Salien- cyi2ploc: saliency-guided image-point cloud localization using contrastive learning,” Information Fusion, vol. 118, p. 103015, 2025

    work page 2025

  3. [3]

    A survey of autonomous driving: Common practices and emerging technologies,

    E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda, “A survey of autonomous driving: Common practices and emerging technologies,” IEEE access, vol. 8, pp. 58 443– 58 469, 2020

    work page 2020

  4. [4]

    Whu-helmet: A helmet-based multisensor slam dataset for the evaluation of real-time 3-d mapping in large-scale gnss-denied environments,

    J. Li, W. Wu, B. Yang, X. Zou, Y . Yang, X. Zhao, and Z. Dong, “Whu-helmet: A helmet-based multisensor slam dataset for the evaluation of real-time 3-d mapping in large-scale gnss-denied environments,” IEEE Trans- actions on Geoscience and Remote Sensing , vol. 61, pp. 1–16, 2023

    work page 2023

  5. [5]

    Data fusion for its: A systematic literature review,

    C. Ounoughi and S. B. Yahia, “Data fusion for its: A systematic literature review,”Information Fusion, vol. 89, pp. 267–291, 2023

    work page 2023

  6. [6]

    Hcto: Optimality-aware lidar inertial odome- try with hybrid continuous time optimization for com- pact wearable mapping system,

    J. Li, S. Yuan, M. Cao, T.-M. Nguyen, K. Cao, and L. Xie, “Hcto: Optimality-aware lidar inertial odome- try with hybrid continuous time optimization for com- pact wearable mapping system,” ISPRS Journal of Pho- togrammetry and Remote Sensing, vol. 211, pp. 228–243, JOURNAL OF LATEX CLASS FILES, VOL. XXX, NO. XXX, XXX XXX 14 2024

    work page 2024

  7. [7]

    Atcm: Aerial-terrestrial lidar-based collaborative simultaneous localization and mapping,

    Y . Xu, C. Chen, B. Yang, W. Wu, S. Sun, Z. Wang, L. Li, and Q. Zou, “Atcm: Aerial-terrestrial lidar-based collaborative simultaneous localization and mapping,” IEEE Transactions on Geoscience and Remote Sensing , 2025

    work page 2025

  8. [8]

    Collaborative sensor network localization: Algorithms and practical issues,

    R. M. Buehrer, H. Wymeersch, and R. M. Vaghefi, “Collaborative sensor network localization: Algorithms and practical issues,” Proceedings of the IEEE , vol. 106, no. 6, pp. 1089–1114, 2018

    work page 2018

Show all 61 references
  1. [9]

    Augmented reality tech- nologies, systems and applications,

    J. Carmigniani, B. Furht, M. Anisetti, P. Ceravolo, E. Damiani, and M. Ivkovic, “Augmented reality tech- nologies, systems and applications,” Multimedia tools and applications, vol. 51, pp. 341–377, 2011

    2011

  2. [10]

    Ua- mpc: Uncertainty-aware model predictive control for mo- torized lidar odometry,

    J. Li, X. Xu, J. Liu, K. Cao, S. Yuan, and L. Xie, “Ua- mpc: Uncertainty-aware model predictive control for mo- torized lidar odometry,” IEEE Robotics and Automation Letters, 2025

    2025

  3. [11]

    Pointnetvlad: Deep point cloud based retrieval for large-scale place recognition,

    M. A. Uy and G. H. Lee, “Pointnetvlad: Deep point cloud based retrieval for large-scale place recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 4470–4479

    2018

  4. [12]

    Minkloc3d: Point cloud based large- scale place recognition,

    J. Komorowski, “Minkloc3d: Point cloud based large- scale place recognition,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2021, pp. 1790–1799

    2021

  5. [13]

    Patchaugnet: Patch feature augmentation-based heterogeneous point cloud place recognition in large-scale street scenes,

    X. Zou, J. Li, Y . Wang, F. Liang, W. Wu, H. Wang, B. Yang, and Z. Dong, “Patchaugnet: Patch feature augmentation-based heterogeneous point cloud place recognition in large-scale street scenes,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 206, pp. 273–292, 2023

    2023

  6. [14]

    Helipr: Heterogeneous lidar dataset for inter-lidar place recognition under spatiotemporal variations,

    M. Jung, W. Yang, D. Lee, H. Gil, G. Kim, and A. Kim, “Helipr: Heterogeneous lidar dataset for inter-lidar place recognition under spatiotemporal variations,” The Inter- national Journal of Robotics Research , vol. 43, no. 12, pp. 1867–1883, 2024

    2024

  7. [15]

    Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges,

    T. Lesort, V . Lomonaco, A. Stoian, D. Maltoni, D. Filliat, and N. D´ıaz-Rodr´ıguez, “Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges,” Information fusion, vol. 58, pp. 52–68, 2020

    2020

  8. [16]

    Expert gate: Lifelong learning with a network of experts,

    R. Aljundi, P. Chakravarty, and T. Tuytelaars, “Expert gate: Lifelong learning with a network of experts,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 3366–3375

    2017

  9. [17]

    Over- coming catastrophic forgetting with hard attention to the task,

    J. Serra, D. Suris, M. Miron, and A. Karatzoglou, “Over- coming catastrophic forgetting with hard attention to the task,” in International conference on machine learning , 2018, pp. 4548–4557

    2018

  10. [18]

    Enhancing few- shot lifelong learning through fusion of cross-domain knowledge,

    Y . Zheng, X. Zhang, Z. Tian, and S. Du, “Enhancing few- shot lifelong learning through fusion of cross-domain knowledge,” Information Fusion , vol. 115, p. 102730, 2025

    2025

  11. [19]

    Incloud: Incremental learning for point cloud place recognition,

    J. Knights, P. Moghadam, M. Ramezani, S. Sridharan, and C. Fookes, “Incloud: Incremental learning for point cloud place recognition,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2022, pp. 8559–8566

    2022

  12. [20]

    Ccl: Continual contrastive learning for lidar place recognition,

    J. Cui and X. Chen, “Ccl: Continual contrastive learning for lidar place recognition,” IEEE Robotics and Automa- tion Letters, vol. 8, no. 8, pp. 4433–4440, 2023

    2023

  13. [21]

    icarl: Incremental classifier and representation learning,

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classifier and representation learning,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , 2017, pp. 2001–2010

    2017

  14. [22]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,” IEEE transactions on pattern analysis and machine intelli- gence, vol. 40, no. 12, pp. 2935–2947, 2017

    2017

  15. [23]

    Oreos: Oriented recognition of 3d point clouds in outdoor scenarios,

    L. Schaupp, M. B ¨urki, R. Dub ´e, R. Siegwart, and C. Ca- dena, “Oreos: Oriented recognition of 3d point clouds in outdoor scenarios,” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2019, pp. 3255–3261

    2019

  16. [24]

    1-day learning, 1- year localization: Long-term lidar localization using scan context image,

    G. Kim, B. Park, and A. Kim, “1-day learning, 1- year localization: Long-term lidar localization using scan context image,” IEEE Robotics and Automation Letters , vol. 4, no. 2, pp. 1948–1955, 2019

    1948

  17. [25]

    Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,

    G. Kim and A. Kim, “Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2018, pp. 4802– 4809

    2018

  18. [26]

    Overlap- net: Loop closing for lidar-based slam,

    X. Chen, T. L ¨abe, A. Milioto, T. R ¨ohling, O. Vysot- ska, A. Haag, J. Behley, and C. Stachniss, “Overlap- net: Loop closing for lidar-based slam,” arXiv preprint arXiv:2105.11344, 2021

    2021 arXiv

  19. [27]

    Improving point cloud based place recognition with ranking-based loss and large batch train- ing,

    J. Komorowski, “Improving point cloud based place recognition with ranking-based loss and large batch train- ing,” in 2022 26th international conference on pattern recognition (ICPR), 2022, pp. 3699–3705

    2022

  20. [28]

    Eca-net: Efficient channel attention for deep convolu- tional neural networks,

    Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca-net: Efficient channel attention for deep convolu- tional neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11 534–11 542

    2020

  21. [29]

    Svt-net: Super light-weight sparse voxel transformer for large scale place recognition,

    Z. Fan, Z. Song, H. Liu, Z. Lu, J. He, and X. Du, “Svt-net: Super light-weight sparse voxel transformer for large scale place recognition,” inProceedings of the AAAI conference on artificial intelligence, vol. 36, no. 1, 2022, pp. 551–560

    2022

  22. [30]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

    2017

  23. [31]

    Netvlad: Cnn architecture for weakly super- vised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: Cnn architecture for weakly super- vised place recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 5297–5307

    2016

  24. [32]

    Pyramid point cloud transformer for large-scale place recogni- tion,

    L. Hui, H. Yang, M. Cheng, J. Xie, and J. Yang, “Pyramid point cloud transformer for large-scale place recogni- tion,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 6098–6107. JOURNAL OF LATEX CLASS FILES, VOL. XXX, NO. XXX, XXX XXX 15

    2021

  25. [33]

    Ndt-transformer: Large-scale 3d point cloud localisation using the normal distribution transform representation,

    Z. Zhou, C. Zhao, D. Adolfsson, S. Su, Y . Gao, T. Duck- ett, and L. Sun, “Ndt-transformer: Large-scale 3d point cloud localisation using the normal distribution transform representation,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 5654– 5660

    2021

  26. [34]

    Look at the whole scene: General point cloud place recognition by classification proxy,

    Y . Xie, B. Wang, H. Wang, F. Liang, W. Zhang, Z. Dong, and B. Yang, “Look at the whole scene: General point cloud place recognition by classification proxy,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 215, pp. 15–30, 2024

    2024

  27. [35]

    A comprehen- sive survey of continual learning: theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehen- sive survey of continual learning: theory, method and application,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 8, pp. 5362–5383, 2024

    2024

  28. [36]

    A contin- ual learning survey: Defying forgetting in classification tasks,

    M. De Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A contin- ual learning survey: Defying forgetting in classification tasks,” IEEE transactions on pattern analysis and ma- chine intelligence, vol. 44, no. 7, pp. 3366–3385, 2021

    2021

  29. [37]

    Large scale incremental learning,

    Y . Wu, Y . Chen, L. Wang, Y . Ye, Z. Liu, Y . Guo, and Y . Fu, “Large scale incremental learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 374–382

    2019

  30. [38]

    Continual learning with deep generative replay,

    H. Shin, J. K. Lee, J. Kim, and J. Kim, “Continual learning with deep generative replay,”Advances in neural information processing systems , vol. 30, 2017

    2017

  31. [39]

    Memory replay gans: Learning to generate new categories without forgetting,

    C. Wu, L. Herranz, X. Liu, J. Van De Weijer, B. Radu- canu et al., “Memory replay gans: Learning to generate new categories without forgetting,” Advances in neural information processing systems , vol. 31, 2018

    2018

  32. [40]

    Some methods for classification and analysis of multivariate observations,

    J. MacQueen, “Some methods for classification and analysis of multivariate observations,” in Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics , vol. 5, 1967, pp. 281–298

    1967

  33. [41]

    Grad-match: Gradient matching based data subset selection for efficient deep model training,

    K. Killamsetty, S. Durga, G. Ramakrishnan, A. De, and R. Iyer, “Grad-match: Gradient matching based data subset selection for efficient deep model training,” in International Conference on Machine Learning , 2021, pp. 5464–5474

    2021

  34. [42]

    Prism: A rich class of parameterized submodular information measures for guided data subset selection,

    S. Kothawade, V . Kaushal, G. Ramakrishnan, J. Bilmes, and R. Iyer, “Prism: A rich class of parameterized submodular information measures for guided data subset selection,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 9, 2022, pp. 10 238– 10 246

    2022

  35. [43]

    Overcoming catas- trophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ra- malho, A. Grabska-Barwinska et al., “Overcoming catas- trophic forgetting in neural networks,” Proceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521...

    2017

  36. [44]

    Packnet: Adding multiple tasks to a single network by iterative pruning,

    A. Mallya and S. Lazebnik, “Packnet: Adding multiple tasks to a single network by iterative pruning,” in Pro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2018, pp. 7765–7773

    2018

  37. [45]

    Pathnet: Evolu- tion channels gradient descent in super neural networks,

    C. Fernando, D. Banarse, C. Blundell, Y . Zwols, D. Ha, A. A. Rusu, A. Pritzel, and D. Wierstra, “Pathnet: Evolu- tion channels gradient descent in super neural networks,” arXiv preprint arXiv:1701.08734 , 2017

    2017 arXiv

  38. [46]

    Bioslam: A bioinspired lifelong memory system for general place recognition,

    P. Yin, A. Abuduweili, S. Zhao, L. Xu, C. Liu, and S. Scherer, “Bioslam: A bioinspired lifelong memory system for general place recognition,” IEEE Transactions on Robotics, vol. 39, no. 6, pp. 4855–4874, 2023

    2023

  39. [47]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neu- big, “Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,” ACM Computing Surveys, vol. 55, no. 9, pp. 1–35, 2023

    2023

  40. [48]

    Parameter-efficient transfer learning for nlp,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” in International conference on machine learning , 2019, pp. 2790–2799

    2019

  41. [49]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  42. [50]

    Continual learning with pre-trained models: A survey,

    D.-W. Zhou, H.-L. Sun, J. Ning, H.-J. Ye, and D.-C. Zhan, “Continual learning with pre-trained models: A survey,” arXiv preprint arXiv:2401.16386 , 2024

    2024 arXiv

  43. [51]

    Learning to prompt for continual learning,

    Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 139–149

    2022

  44. [52]

    Dualprompt: Complementary prompting for rehearsal-free continual learning,

    Z. Wang, Z. Zhang, S. Ebrahimi, R. Sun, H. Zhang, C.-Y . Lee, X. Ren, G. Su, V . Perot, J. Dy et al., “Dualprompt: Complementary prompting for rehearsal-free continual learning,” in European Conference on Computer Vision , 2022, pp. 631–648

    2022

  45. [53]

    S-prompts learn- ing with pre-trained transformers: An occam’s razor for domain incremental learning,

    Y . Wang, Z. Huang, and X. Hong, “S-prompts learn- ing with pre-trained transformers: An occam’s razor for domain incremental learning,” Advances in Neural Information Processing Systems, vol. 35, pp. 5682–5695, 2022

    2022

  46. [54]

    Generating instance-level prompts for rehearsal-free continual learn- ing,

    D. Jung, D. Han, J. Bang, and H. Song, “Generating instance-level prompts for rehearsal-free continual learn- ing,” in Proceedings of the IEEE/CVF International Con- ference on Computer Vision , 2023, pp. 11 847–11 857

    2023

  47. [55]

    Efficient svm training using low-rank kernel representations,

    S. Fine and K. Scheinberg, “Efficient svm training using low-rank kernel representations,” Journal of Machine Learning Research, vol. 2, no. Dec, pp. 243–264, 2001

    2001

  48. [56]

    Blip-2: Boot- strapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Boot- strapping language-image pre-training with frozen image encoders and large language models,” in International conference on machine learning , 2023, pp. 19 730– 19 742

    2023

  49. [57]

    1 year, 1000 km: The oxford robotcar dataset,

    W. Maddern, G. Pascoe, C. Linegar, and P. Newman, “1 year, 1000 km: The oxford robotcar dataset,” The International Journal of Robotics Research , vol. 36, no. 1, pp. 3–15, 2017

    2017

  50. [58]

    Mul- ran: Multimodal range dataset for urban place recogni- tion,

    G. Kim, Y . S. Park, Y . Cho, J. Jeong, and A. Kim, “Mul- ran: Multimodal range dataset for urban place recogni- tion,” in 2020 IEEE international conference on robotics JOURNAL OF LATEX CLASS FILES, VOL. XXX, NO. XXX, XXX XXX 16 and automation (ICRA) , 2020, pp. 6246–6253

    2020

  51. [59]

    Domain generalization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain generalization: A survey,” IEEE transactions on pattern analysis and machine intelligence , vol. 45, no. 4, pp. 4396–4415, 2022

    2022

  52. [60]

    Self-supervised learning: Generative or contrastive,

    X. Liu, F. Zhang, Z. Hou, L. Mian, Z. Wang, J. Zhang, and J. Tang, “Self-supervised learning: Generative or contrastive,” IEEE transactions on knowledge and data engineering, vol. 35, no. 1, pp. 857–876, 2021

    2021

  53. [61]

    Gra- dient based sample selection for online continual learn- ing,

    R. Aljundi, M. Lin, B. Goujaud, and Y . Bengio, “Gra- dient based sample selection for online continual learn- ing,” Advances in neural information processing systems, vol. 32, 2019. BIOGRAPHY SECTION Xianghong Zou received the B.S. degree in geo- matics engineering, the M.S. ...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.