acceptodds
Under review as a conference paper at ICLR 2027

Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection

Abstract

Trajectory anomaly detection underpins applications from fraud detection to urban mobility analysis. Dense GPS preserves fine-grained evidence such as abnormal speeds and short-duration events, but its quadratic cost makes multi-month analysis intractable; sparse stay-point methods scale by discarding that evidence and require a separate modeling regime. We argue that this bottleneck is unnecessary: dense and sparse trajectories share a natural two-dimensional cyclic structure along within-day and across-day axes. TITAnD (Trajectory Image Transformer for Anomaly Detection) is the first framework to cast both in a single representation, a Hyperspectral Trajectory Image (HTI), a day × time-of-day grid whose channels encode spatial, semantic, temporal, and kinematic information. Under this formulation, agent-level detection reduces to image classification and temporal localization to semantic segmentation. The Cyclic Factorized Transformer (CFT) models the two temporal axes directly, reducing attention cost by up to two orders of magnitude. With the HTI, multi-month dense anomaly detection becomes feasible for the first time, and CFT keeps it accurate and fast. Empirically, TITAnD matches or beats every evaluated baseline in AUC-PR on all four benchmarks. It surpasses vision models such as U-Net while using as few as 6.5M parameters, and it runs 11–75× faster than a capacity-matched flat Transformer on the same trajectory images.

Primary area:
learning on time series and dynamical systems
Keywords:
trajectory anomaly detection, human mobility, long-context modeling, hyperspectral trajectory image, cyclic factorized transformer, semantic segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WxxMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper proposes TITAnD, a supervised framework for multi-month trajectory anomaly detection that unifies dense GPS trajectories and sparse stay-point sequences through a Hyperspectral Trajectory Image (HTI) representation. The key idea is to reorganize long-term trajectories into a two-dimensional day × time-of-day grid, where each cell encodes spatial, semantic, temporal, and kinematic information. To efficiently model this representation, the authors introduce a Cyclic Factorized Transformer (CFT), which alternates intra-day and inter-day attention to capture within-day movement patterns and recurring behavioral routines across days, while reducing the computational complexity of standard self-attention.

Critical strengths

S1. The proposed HTI representation provides a unified framework for detecting anomalies in both dense GPS trajectories and sparse stay-point sequences.

S2. The CFT effectively captures intra-day and inter-day behavioral patterns while significantly reducing computational costs compared to standard Transformers.

S3. TITAnD achieves competitive performance across four benchmarks.

Critical weaknesses

W1. The proposed HTI representation and CFT architecture share substantial similarities with existing period-based time-series models and axial attention mechanisms. Although these connections are discussed, the paper lacks direct experimental comparisons with closely related methods, making it difficult to establish the novelty and distinct benefits of the proposed approach.

W2. Lack of evaluation on real-world trajectories. All four benchmarks rely on simulated trajectories or synthetically injected anomalies.

W3. The paper reports training and validation splits but does not clearly establish a test set. The appendix states that the reported results use checkpoints selected by validation performance.

W4. The superiority of full CFT over single-axis attention does not establish that the specific cyclic alignment is essential, as opposed to the general benefits of factorized attention.

W5. The proposed HTI is essentially a structured day x time-of-day feature tensor rather than an image in the conventional sense. Referring to it as a "hyperspectral image" may be misleading, as its channels represent learned trajectory features rather than spectral information. The authors should better justify this terminology.

W6. Insufficient comparisons with trajectory anomaly detection baselines, which are widely studied in the community. In particular, the paper explicitly distinguishes dense and sparse trajectory anomaly detection, the trajectory-specific baselines considered are limited to sparse methods.

W7. The experimental section is not well structured, making it difficult to follow the evaluation protocol and interpret the results. A clearer organization and more systematic analysis would improve readability and strengthen the experimental discussion.

Questions for authors

NA

5 comments
AuthorsOpenReview ↗

Reply to R-WxxM (Part 1 of 1)

Reply to R-WxxM

We thank the reviewer for the detailed feedback. The concerns about novelty and comparisons with closely related methods (W1), simulated data (W2), the independent test set (W3), and cyclic alignment (W4) are addressed in the General Response, Sections 2–3, 5, 1, and 4, respectively. Below, we address the remaining concerns.

1. Dense trajectory baselines

W6: "Insufficient comparisons with trajectory anomaly detection baselines."

To address this concern, we added a supervised adaptation of GM-VSAE [1], a dense trajectory anomaly detection method, using its released trip encoder. Existing dense trajectory anomaly detection methods, such as iBAT, GM-VSAE, DeepTEA [2], and FOTraj, primarily focus on detecting anomalies within individual trips. Our task differs as we detect and localize anomalies over an agent's multi-month trajectory, including behavior outside trips. In Dense Tokyo, only 16–18% of anomalous 5-minute slots occur within trips.

We evaluated GM-VSAE in two configurations: its original trip-only formulation and an extension that also considers stay-points. Both were trained using the same supervision as the other baselines. Further implementation details are provided in the supplementary material.

Dense Tokyo, TestTemporal AUC-PRAgent AUC-PR
CFT (ours)0.831.00
GM-VSAE [1] (trip-only adaptation)0.110.79
GM-VSAE [1] (trips + stay-points)0.040.61

All methods are evaluated on the same 15-minute slots for a fair comparison.

GM-VSAE achieves meaningful agent-level detection performance (0.79 AUC-PR), but struggles with temporal localization. This reflects a limitation of trip-based methods when applied to complete multi-month behavioral histories, where many anomalies occur outside individual trips. CFT, in contrast, localizes anomalous behavior across the full trajectory.

2. The term "hyperspectral"

W5: "Referring to it as a 'hyperspectral image' may be misleading."

We use "hyperspectral" as an analogy to the multichannel image structure of HTI, rather than to imply that its channels represent optical spectra. Each pixel encodes learned spatial, semantic, temporal, and kinematic features. We will clarify this terminology in Sec. 3. If the reviewer still finds the term inappropriate, we are open to changing it and would appreciate any suggestions.

3. Structure of the experimental section

W7: "The experimental section is not well structured, making it difficult to follow the evaluation protocol and interpret the results."

We appreciate the reviewer's feedback on the experimental organization. In the revised paper, Sec. 4.1 now gives the train, validation, and test sizes of every benchmark and states that all metrics are computed only over the evaluation period; Table 3 reports the held-out Test results; and each ablation in Sec. 4.3 states the question it answers before its result. We would be glad to clarify any further experiments, tables, or explanations the reviewer found unclear.

References

[1] Y. Liu et al. Online Anomalous Trajectory Detection with Deep Generative Sequence Modeling. ICDE 2020.

[2] X. Han et al. DeepTEA: Effective and Efficient Online Time-dependent Trajectory Outlier Detection. PVLDB 2022.

Thank you for the detailed rebuttal. I appreciate the efforts made to strengthen the paper. These new results and clarifications have addressed several of my major concerns and improved my confidence in the technical soundness of the work. I have therefore raised my Soundness score. I have one additional suggestion regarding the revised manuscript: it would be helpful if the authors could highlight the changes made in response to the reviewers' comments, so that reviewers can more easily identify and assess the revisions.

We thank the reviewer for the positive follow-up and are glad that the additional experiments and clarifications have addressed several of the major concerns.

Following the suggestion, we have uploaded a revised manuscript with all changes highlighted in blue. The main revisions are:

  • Held-out Test evaluation for every benchmark, with disjoint Train, Validation, and Test agents (Sec. 4.1, Table 3, Appendix H).
  • Comparisons with TimeMixer++ and GM-VSAE (Sec. 4.2, Appendix I).
  • Distinction from axial attention (Sec. 3.3, Appendix G).
  • Calendar-alignment, history-length, and label-efficiency studies (Sec. 4.3, Appendix J).
  • Detection of unseen anomaly groups (Appendix K).
  • Results on the public SF-TPAD benchmark (Appendix L).

We appreciate the reviewer's constructive feedback, which has helped strengthen the paper

Reviewer_Ca8VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TITAnD, an end-to-end supervised framework for agent-level anomaly detection and temporal localization over multi-month trajectories. It supports dense GPS streams and sparse stay-point logs through data-specific encoders.

Both encoders produce a Hyperspectral Trajectory Image (HTI), a day × time-of-day grid with spatial, semantic, temporal, and kinematic channels. The Cyclic Factorized Transformer (CFT) processes this grid using within-day and across-day attention. This structure captures daily episodes and recurring routines at a lower attention cost than a flat Transformer.

The authors compare TITAnD with baselines on four dense and sparse benchmarks. They also examine the input features and attention axes through ablation studies and measure computational efficiency. TITAnD matches or exceeds the evaluated baselines in the reported AUC-PR results, with larger gains in temporal localization than in agent-level detection. It also achieves faster inference than a capacity-matched flat Transformer using the same trajectory images.

Critical strengths
  1. The method is well motivated by the within-day and across-day structure of human mobility, and it improves temporal localization. CFT models both within-day episodes and recurring routines across days. In Table 1, temporal AUC-PR increases from 0.69 for the strongest evaluated baseline to 0.84 on Dense Tokyo, and from 0.46 to 0.59 on NumoSim-LA. These gains support the method’s practical value in identifying when anomalous behavior occurs, beyond detecting anomalous agents.

  2. The method processes long trajectories efficiently, with ablations supporting the use of both attention axes. Figure 4 shows that CFT runs 11 to 75 times faster than a capacity-matched flat Transformer on sparse stay-point HTIs over horizons of 2 to 12 months. Table 2 also shows that combining intra-day and inter-day attention outperforms either axis alone, with the single-axis models adjusted to match the parameter count. The efficiency measurements and ablations provide evidence that the factorized design can reduce computational cost while retaining the interactions needed for anomaly localization.

Critical weaknesses
  1. The contribution beyond the closest existing methods is not sufficiently established. Prior work has explored both reshaping time series into two-dimensional representations and applying attention along separate axes. The paper compares CFT with a flat Transformer and several vision backbones, but lacks direct comparisons with the closest period-folding or axial-attention methods. The reported results support the proposed framework over the selected baselines. However, they leave unclear how much the trajectory-specific design adds over similar existing architectures. This limits my assessment of the method’s novelty and additional value.

  2. The contribution of multi-month history to detection performance is not yet clear. Multi-month trajectory modeling is a central contribution of the paper. Figure 4 reports computational efficiency over different observation horizons, but does not show how detection and temporal localization performance change with history length. The ablations in Table 2 support combining intra-day and inter-day attention, yet do not establish whether shorter histories would already be sufficient to achieve the reported results. A comparison using different history lengths, with the evaluation period and other experimental conditions fixed, would help establish whether the model benefits from longer-term behavioral patterns and how large that benefit is. Without this evidence, it is difficult to assess the additional detection value of multi-month modeling.

Questions for authors
  1. Are the results in Table 1 computed on an independent test set? Please clarify which data are used for training, checkpoint selection, and final evaluation. For mIoU, the paper states that the threshold maximizes F1 on the evaluated predictions. Could you also report results with thresholds selected only on the validation set and then applied to an independent test set?

  2. How does CFT differ from standard axial attention and existing period-folding time-series models? Could you compare it with one of the closest alternatives on a main benchmark, using the same input features and a comparable training budget? This would help establish the benefit of the specific design choices in CFT.

  3. Could you compare anomaly detection and temporal localization performance across different history lengths, keeping the evaluated agents, evaluation period, and other experimental conditions fixed? Figure 4 shows the computational efficiency of processing long sequences. I would also like to know whether longer histories improve detection performance, or whether shorter histories achieve similar results. This experiment would directly test the additional benefit of multi-month trajectory modeling.

I would be willing to raise my score if the authors’ responses and additional experiments resolve these concerns.

5 comments
AuthorsOpenReview ↗

Reply to R-Ca8V (Part 1 of 1)

Reply to R-Ca8V

We thank the reviewer for the detailed feedback and suggestions. The concerns about the independent test set (Q1) are addressed in General Response Section 1, while the comparisons with closely related methods (W1, Q2), including TimeMixer++ and the distinction from standard axial attention, are addressed in Sections 2–3. Our calendar-alignment experiment in Section 4 further examines the importance of CFT's cyclic inductive bias. Below, we address the remaining concern about history length.

1. History length

W2, Q3: "Could you compare anomaly detection and temporal localization performance across different history lengths, keeping the evaluated agents, evaluation period, and other experimental conditions fixed?"

We appreciate the suggestion and conducted an additional experiment to investigate how much historical context contributes to anomaly detection. We train CFT with different history lengths while keeping the evaluated agents, evaluation period, labels, and training configuration fixed. We use NumoSim-LA, the largest public benchmark, to make the experiment reproducible.

Historical contextTemporal AUC-PR (Test)Agent AUC-PR (Test)
0 days0.410.53
7 days0.580.65
14 days0.570.66
28 days (full)0.580.66

All configurations use the same training settings, so the full-history result may differ slightly from the main comparison.

The results show that historical context substantially improves anomaly detection, with temporal AUC-PR increasing from 0.41 to 0.58 after adding seven days of history. Beyond seven days, performance remains largely unchanged. We hypothesize that the amount of history needed depends on the agents' patterns of life and the anomaly types in each dataset. For NumoSim-LA, one week of history appears sufficient to capture most of the relevant behavioral patterns.

Importantly, the amount of historical context needed for detection and the ability to process multi-month trajectories are two different questions. HTI enables multi-month dense trajectory modeling, while CFT exploits its cyclic structure to improve detection performance and computational efficiency. As shown in Fig. 4, CFT is 11–75× faster than a similarly sized flat Transformer across 2–12 months. Further implementation and efficiency details are provided in the paper and supplementary material.

I appreciate the authors' efforts in conducting additional experiments and revising the paper. The inclusion of evaluations on an independent test set, comparisons with TimeMixer++ , and experiments regarding historical sequence length have significantly strengthened the response to my initial concerns.

Regarding historical sequence length, the new experiments demonstrate that incorporating historical information yields clear benefits; however, on the NumoSim-LA dataset, these gains plateau after seven days. I accept the authors' distinction between "efficiently processing multi-month trajectories" and "deriving additional benefits from longer historical sequences" this point has been adequately addressed.

As for the evaluation protocol, the newly added independent test results have alleviated my primary concerns. That said, Appendix H notes that Table 1 still relies on the original model and threshold selection process. I suggest explicitly stating this in the description for Table 1, and using the independent test results presented in Table 3 and Appendix Table 7 as the primary basis for the paper's main conclusions.

The additional comparison with TimeMixer++ is very helpful. I would like to clarify the training configurations: Table 5 indicates 40 training epochs for the Dense Tokyo dataset, whereas Appendix I specifies 20 epochs for TimeMixer++ . Does the CFT model in Table 8 follow the former configuration? Furthermore, has TimeMixer++ fully converged within the given computational budget? Please provide a brief explanation regarding the choice of budget or the model's performance on the validation set, so that this comparison can be properly assessed.

We thank the reviewer for the positive assessment and constructive follow up feedback. We are glad that the independent test evaluation, TimeMixer++ comparison, and history-length experiments have addressed the main concerns. We clarify the two remaining points below.

  1. Evaluation protocol: We agree with the suggestion. We will clarify in the Table 1 caption that it reports the original evaluation protocol, where checkpoints and thresholds were selected on the evaluated agents. The main conclusions will instead be based on the independent held-out Test results in Table 3 and Appendix Table 7.

  2. TimeMixer++ Training: Yes, the CFT results in Table 8 follow the training configuration in Table 5. For TimeMixer++, validation temporal AUC-PR peaked at 0.66 at epoch 7 before declining to 0.59 at epoch 20, while training loss continued decreasing, suggesting that training had converged. To further verify this, we retrain the model doubling the training epochs. Validation performance did not improve, and the checkpoint selected using the best validation score also showed no improvement on the held out Test set. These results provide strong evidence that the performance gap is not due to insufficient training.

Finally, we appreciate the reviewer's constructive suggestions, which have helped improve the paper.

Reviewer_xzE5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses supervised anomaly detection in multi-month human trajectories. It introduces a strong prior that exploits the periodic and cyclical patterns inherent in human mobility, such as daily commutes, to improve trajectory anomaly detection in urban mobility analysis. The paper formulate trajectory anomaly detection as a computer vision task by mapping GPS data onto a day × time-of-day grid, where each pixel represents mobility behaviour within a fixed time window. Anomalies are localised through supervised semantic segmentation. The proposed method achieves strong performance on synthetic data (4 benchmarks). On each evaluation, it is on paar with adapted sparse baselines and other backbones.

Critical strengths
  • It creates a compact representation for months of GPS data. This makes long-horizon dense GPS detection feasible.
  • The approach maps both sparse stay-points and dense GPS streams into the same representation.
  • It shows strong inductive biases based on how human routines work can strongly improve performance i.e., modelling cyclic patterns via intra-day and inter-day attention.
  • It detects both timing and location anomalies at similar rates (App. E).
  • It is fast and lightweight compared with a flat Transformer (11–75× faster, 6.5M parameters on sparse data), although U-Net and SegFormer are faster.
  • The paper is well-written and easy to follow.
Critical weaknesses
  • Evaluation protocol. The splits contain only training and validation agents (Table 3), and Table 1 reports the checkpoint with the best validation score (App. D). Model selection and reporting therefore use the same data. The mIoU threshold is also chosen on the evaluated predictions, and no standard deviations are given. With 30 anomalous validation agents on PoL, 0.71 against 0.68 for SegFormer is likely within noise.
  • Supervised training with synthetic labels. The paper mentions itself that verified real-world mobility anomaly labels are scarce. Supervised anomaly detection is probably not the way to scale this. Also real-world data is likely more messy, making it harder to detect anomalies.
  • The approach doesn’t test transfer to anomaly types unseen in training. The model could just learn the injection process.
  • The reported baselines were designed around self-supervised objectives, but here they are trained in a supervised fashion. They are potentially favouring the proposed method in the evaluations (USTAD and ICAD reach 0.06 agent AUC-PR on PoL, chance at a 5% positive rate). Since all methods share the same labels, the comparison isolates the design only within the supervised setting. It leaves open whether HTI and CFT help without labels.
  • Dense results are obtained on a private dataset that cannot be reproduced without access.
  • TimeMixer++ is only discussed, but not actually benchmarked. TimeMixer++ (ICLR 2025) already applies dual-axis attention to time images, so it is the closest design to CFT and the comparison that would show what CFT adds.
  • The used time grid seems very rigid, i.e., human behaviour has to align with it to be modelled correctly (Calendar days and fixed slots).
Questions for authors
  • Do HTI and CFT also help without labels? USTAD, ICAD, and LM-TAD were run as supervised variants with their self-supervised heads replaced. Could you report them in their original unsupervised form, and train a self-supervised TITAnD (for example, masked-slot reconstruction on the HTI)? That would show whether the representation helps without labels, which is how most of the related work operates.
  • Does the model generalise to novel unseen anomaly types? For example, train on unfamiliar-location anomalies and test on shifted-time revisits in NumoSim-LA (App. E).
  • Since inter-day attention compares only the same slot across days, how sensitive is the method to the fixed day and slot grid, e.g., shifted day boundary, night-shift workers or activities that span midnight?
  • Will the dense data or an equivalent public dense benchmark be released?
  • Is there a test split separate from the validation split used for checkpoint selection?
  • Are the temporal metrics computed over all slots, including the anomaly-free baseline days (e.g. days 0–32 on Dense Tokyo, App. F)? Day position alone could then raise AUC-PR, so please also report metrics over the evaluation window only.
6 comments
AuthorsOpenReview ↗

Reply to R-xzE5 (Part 1 of 3)

Reply to R-xzE5

We thank the reviewer for the detailed feedback. The concerns about the held-out test set (W1, Q5), TimeMixer++ (W6), cyclic alignment (W7), and synthetic data (W2) are addressed in the General Response, Sections 1, 2, 4, and 5, respectively. Below, we address the remaining concerns.

1. PoL and the spread of results

W1: "With 30 anomalous validation agents on PoL, 0.71 against 0.68 for SegFormer is likely within noise."

We appreciate the reviewer's concern and agree that the difference on PoL should be interpreted carefully. PoL has 2,400 training agents and 15 anomalous agents in the new held-out Test set. To better understand the variability, we increased the number of training seeds from six to nine and report the mean and standard deviation.

MethodAgent AUC-PR (mean ± SD, 9 seeds)
CFT (ours)0.649 ± 0.033
SegFormer0.595 ± 0.227

CFT achieves a higher mean AUC-PR with substantially lower variability across seeds. We hypothesize that PoL's limited training data may contribute to the observed variability. Given 15 anomalous Test agents, we treat PoL as supporting evidence; our main conclusions rest on the other three benchmarks, where CFT's temporal margins are larger (General Response, Section 1).

2. Anomaly types unseen in training

W3, Q2: "The model could just learn the injection process."

We appreciate the reviewer's concern and conducted additional experiments to evaluate whether CFT generalizes to anomaly types unseen during training. The reviewer suggests training on unfamiliar-location anomalies and testing on shifted-time revisits in NumoSim-LA. However, these two anomaly types come from the same injection process. In particular, 98% of shifted-time revisits occur adjacent to inserted stays, and 237 of 238 anomalous training agents contain both types. Thus, training on one type would still expose the model to the other.

We therefore use Dense Tokyo, where distinct anomaly groups injected by an independent red team affect different agents, allowing us to hold out entire groups from training and validation. The red team withheld the injection rules from us (Appendix E). We train using only one anomaly group and evaluate separately on agents with the seen group (252) and unseen groups (238), using the same normal agents for comparison.

Dense Tokyo, temporal AUC-PR (Test)Seen typesUnseen types
CFT (ours)0.950.24
Chronos-20.720.07

Both methods are evaluated on the same 15-minute slots of the held-out Test set.

Although performance drops on unseen anomaly types, CFT retains about a quarter of its seen-type AUC-PR (0.95 to 0.24), compared with less than a tenth for Chronos-2 (0.72 to 0.07). Importantly, unseen anomalies account for only 0.38% of the evaluated slots, making CFT's 0.24 AUC-PR approximately 62× chance level, compared with about 18× for Chronos-2. The unseen anomaly groups also differ in duration from the training group, ranging from under an hour to several hours, which makes accurate temporal localization more challenging.

We observe a similar advantage on Dense2Sparse, where CFT achieves 0.12 temporal AUC-PR compared with 0.04 for DINOv2+Adapter on unseen anomaly types. These results show that CFT's performance is not limited to recognizing anomaly types seen during training. We include further experimental details and analyses in the supplementary material.

AuthorsOpenReview ↗

Reply to R-xzE5 (Part 2 of 3)

3. Supervised baselines

W4: "… designed around self-supervised objectives, but here they are trained in a supervised fashion … potentially favouring the proposed method … (USTAD and ICAD reach 0.06 agent AUC-PR on PoL …)."

Q1: "Do HTI and CFT also help without labels?"

We appreciate the suggestion. Our work focuses on supervised trajectory anomaly detection, which remains relatively underexplored compared with the extensive literature on self-supervised anomaly detection. We do not claim that TITAnD outperforms existing methods in their original self-supervised setting.

For a fair comparison within the supervised setting, we train USTAD, ICAD, and LM-TAD end-to-end using the same labels, agents, and model-selection protocol. These adapted baselines achieve competitive results: on NumoSim-LA, USTAD reaches 0.53 agent AUC-PR, close to U-Net (0.57), while ICAD reaches 0.30.

PoL is more challenging for these sequence baselines. It provides agent-level rather than stay-level labels, with only about 120 anomalous agents among 2,400 training agents. Moreover, anomalies occur only in the final 14 days of a 464-day trajectory. These limited and coarse labels make learning anomaly localization difficult for stay-sequence models, which may help explain why USTAD and ICAD remain near chance (0.06 agent AUC-PR against a 0.05 chance level).

More importantly, our conclusions do not depend solely on these adapted baselines. We also compare CFT with a flat Transformer, U-Net, SegFormer, and DINOv2+Adapter using the same HTI inputs. These comparisons independently support the benefit of CFT's cyclic inductive bias. Evaluating HTI and CFT under self-supervised learning is a natural next step, and we will discuss this direction in the limitations and future work.

4. How many labels are needed?

W2: "Supervised anomaly detection is probably not the way to scale this."

To address this concern, we conducted an additional experiment to investigate whether CFT can learn effectively with fewer anomaly labels. We use NumoSim-LA, the larger public benchmark with temporal anomaly annotations, and compare CFT against USTAD, the strongest competing sequence baseline for temporal localization. We train both methods using 25%, 50%, and 100% of anomalous training agents while keeping the Validation and Test sets fixed.

Anomalous training agentsCFT temporal / agent AUC-PRUSTAD temporal / agent AUC-PR
25%0.38 / 0.490.38 / 0.50
50%0.48 / 0.600.43 / 0.50
100%0.58 / 0.660.43 / 0.49

With only 50% of the anomalous training agents, CFT already outperforms USTAD trained with 100% (0.48 vs. 0.43 temporal and 0.60 vs. 0.49 agent AUC-PR). With 25%, the methods perform similarly. These results suggest that CFT can make effective use of limited anomaly labels, supporting supervised detection even when labeled anomalies are scarce.

5. Evaluation window only

Q6: "Are the temporal metrics computed over all slots, including the anomaly-free baseline days?"

All temporal AUC-PR results in the original paper are computed exclusively over the evaluation period, excluding the historical baseline. The historical period provides context for detecting future anomalies but does not contribute to the evaluation metrics. Thus, the reported results are not inflated by the anomaly-free baseline days. We will clarify this in Sec. 4.1.

AuthorsOpenReview ↗

Reply to R-xzE5 (Part 3 of 3)

6. Evaluation on a public dense trajectory benchmark

W5, Q4: "Dense results are obtained on a private dataset that cannot be reproduced without access."

We acknowledge the reviewer's concern. While Dense Tokyo and Dense2Sparse belong to an external provider and cannot be released by us, we have extended our evaluation to SF-TPAD [1], a publicly available dense GPS anomaly benchmark built upon SF-LIFE [2]. SF-TPAD provides anomalous and normal agents with 5-second GPS trajectories spanning 70 days. Both CFT and Chronos-2 were trained from scratch using San Francisco-specific spatial and semantic features. We use an agent-disjoint train/validation/test split of 120/40/40 agents; the dataset and split details are in the supplementary material. Although SF-TPAD has relatively few agents, it provides temporal annotations for anomalous stays, yielding approximately 10,000 anomalous 5-minute slots for training.

SF-TPAD (3 seeds)Temporal AUC-PRAgent AUC-PR
CFT (ours)0.407 ± 0.0450.851 ± 0.060
Chronos-20.257 ± 0.0340.779 ± 0.058

Both methods are evaluated on the same Test slots, with chance levels of approximately 0.004 temporal and 0.50 agent AUC-PR. CFT outperforms Chronos-2 by 0.15 temporal AUC-PR (approximately 110× vs. 70× chance), consistent with its advantage on Dense Tokyo. These results extend our dense trajectory evaluation beyond the private Tokyo data to a separately developed, publicly available benchmark.

Given the limited rebuttal period, we prioritized Chronos-2, the strongest competing baseline for dense temporal localization in our original experiments. We will extend the comparison to the remaining baselines in the camera-ready version. Furthermore, two of our original four benchmarks, NumoSim-LA and PoL, are already public. We will also release our implementation, model weights, training configurations, and evaluation code to support further research.

7. The fixed day-and-slot grid

Q3: "How sensitive is the method to the fixed day and slot grid, e.g., shifted day boundary, night-shift workers or activities that span midnight?"

CFT models recurring behavior at the same clock time across days, regardless of conventional working hours. Thus, a night-shift worker who regularly works at 2 am still exhibits a consistent pattern in the HTI, while activities spanning midnight are represented across adjacent daily rows.

We also examined sensitivity to the day boundary on NumoSim-LA, where the original HTI uses UTC midnight rather than local midnight. Shifting the boundary to local midnight has minimal impact on performance, with only a 0.04 decrease in temporal AUC-PR and a 0.03 decrease in agent AUC-PR. This shows that CFT is relatively insensitive to where the day begins, as long as the 24-hour cycle is preserved. In contrast, disrupting the cycle itself substantially reduces performance, as shown in General Response, Section 4.

References

[1] Y. Liu et al. Synthesizing Human Trajectory Anomalies: An LLM-Driven Framework with Map-Constrained Routing and Sensor Noise Simulation. arXiv:2606.10314, 2026.

[2] Algama et al. SF-LIFE. arXiv:2606.00430, 2026.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response (Part 3 of 3)

4. Is the calendar alignment essential?

R-WxxM (W4): "... does not establish that the specific cyclic alignment is essential, as opposed to the general benefits of factorized attention."

R-xzE5 (W7): "The used time grid seems very rigid ..."

We thank the reviewers for raising this concern. To investigate whether the performance gains come from cyclic alignment rather than factorized attention alone, we conducted an additional experiment where we deliberately disrupted the cyclic alignment of the HTI (24 hours for intra-day attention) while keeping the CFT architecture unchanged. We use NumoSim-LA, a large public benchmark, to make the experiment reproducible.

Intra-day attention spanTemporal AUC-PR (Test)Agent AUC-PR (Test)
24 hours (daily cycle)0.580.66
23 hours0.420.59
25 hours0.430.59

All configurations use the same training settings, so the 24-hour result may differ slightly from the main comparison. With 23- or 25-hour spans, the columns gradually drift away from the same clock time across days, even though the model still uses factorized attention. As a result, temporal AUC-PR drops from 0.58 to 0.42–0.43, while agent AUC-PR drops from 0.66 to 0.59. These results show that the improvement is not simply due to factorized attention, but also to aligning the attention axes with the cyclic structure of human mobility.

5. Simulated data and synthetic anomalies

R-xzE5 (W2): "Supervised training with synthetic labels."

R-WxxM (W2): "Lack of evaluation on real-world trajectories."

We acknowledge that evaluation on naturally occurring anomalies remains one of the limitations, as discussed in Sec. 5. However, we would like to highlight that verified anomaly labels for long-term human mobility are extremely difficult to obtain. Importantly, the research community has widely adopted simulated data when real-world data or labels are unavailable, for instance in CARLA [1] for autonomous driving, Habitat [2] for embodied AI, and UBnormal [3] for video anomaly detection.

Notably, the benchmarks used in our paper are indeed grounded in realistic human mobility patterns rather than arbitrary anomaly injections. For example, NumoSim incorporates mobility patterns derived from real-world travel data, while Patterns of Life simulates interpretable behavioral anomalies such as skipping work or changing daily activities. Our qualitative analysis (Appendix F) further supports this, as we also observed anomalies involving visits to unusual places at unusual times, reflecting recognizable real-world behavioral deviations. We believe these benchmarks provide meaningful controlled evaluations, while recognizing that generalization to naturally occurring anomalies remains an open question.

Revisions to the paper. We have uploaded a revised paper: held-out Test evaluation and split sizes (Sec. 4.1, Table 3, App. H); TimeMixer++ and GM-VSAE (Sec. 4.2, App. I); how CFT differs from axial attention (Sec. 3.3, App. G); calendar alignment, history length and label efficiency (Sec. 4.3, App. J); unseen anomaly groups (App. K); SF-TPAD (Sec. 4.1-4.2, App. L); and clarifications on the term "hyperspectral" (Sec. 3.1), the evaluation period (Sec. 4.1), and self-supervised learning as future work (Sec. 5).

References

[1] A. Dosovitskiy et al. CARLA: An Open Urban Driving Simulator. CoRL 2017.

[2] M. Savva et al. Habitat: A Platform for Embodied AI Research. ICCV 2019.

[3] A. Acsintoae et al. UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection. CVPR 2022.


AuthorsOpenReview ↗

General response (Part 2 of 3)

2. Comparison with the closest methods

R-xzE5 (W6): "TimeMixer++ is only discussed, but not actually benchmarked."

R-Ca8V (Q2): "Could you compare it with one of the closest alternatives on a main benchmark, using the same input features and a comparable training budget?"

R-WxxM (W1): "... the paper lacks direct experimental comparisons with closely related methods ..."

We thank the reviewers for the suggestion. We trained TimeMixer++, the closest related method, using the same spatial, semantic, temporal, and kinematic features as CFT. Our results show that CFT substantially outperforms TimeMixer++ in temporal anomaly localization.

TimeMixer++ is the closest design to ours, as R-xzE5 points out. It folds a time series into a 2D grid based on detected periods and then applies attention along both axes. R-Ca8V and R-WxxM also ask for comparisons with closely related period-folding or axial-attention methods, so we use TimeMixer++ to address the concerns of all three reviewers.

Notably, TimeMixer++ expects a regularly sampled 1D time series. We therefore evaluate it on Dense Tokyo, providing the trajectory as a regular time series with the same underlying features CFT uses: position, nearby place categories, movement, and calendar information. We trained it on the same data and with the same training budget as Chronos-2, with details of the training recipe in the supplementary material. All methods were evaluated on the held-out Test set, with results shown below. All methods are scored on CFT's 15-minute slots; for the 5-minute models, each slot takes the maximum of its three 5-minute scores.

MethodTemporal AUC-PRAgent AUC-PR
CFT (ours)0.831.00
TimeMixer++0.651.00
Chronos-20.700.86

These results demonstrate that the benefit of TITAnD goes beyond transforming trajectories from 1D to 2D or applying axial attention. Despite receiving the same underlying features, TimeMixer++ achieves 0.65 temporal AUC-PR compared with 0.83 for CFT. This 0.18 gap highlights the importance of organizing trajectories around the natural within-day and across-day structure of human mobility, rather than relying on generic period folding and factorized attention. Our calendar-alignment experiment, discussed below, further supports this: changing the row length from 24 hours to 23 or 25 hours reduces temporal AUC-PR from 0.58 to 0.42–0.43.

3. How does CFT differ from standard axial attention?

R-Ca8V (Q2): "How does CFT differ from standard axial attention and existing period-folding time-series models?"

R-WxxM (W1): "... share substantial similarities with existing period-based time-series models and axial attention mechanisms."

R-WxxM (W4): "... does not establish that the specific cyclic alignment is essential, as opposed to the general benefits of factorized attention."

CFT differs from existing axial-attention methods in its primary objective and the meaning of its attention axes. While Axial Transformer and Axial-DeepLab primarily focus on reducing the computational cost of long-range attention, CFT is designed to exploit the cyclic structure of human mobility. The key differences are summarized below.

Axial Transformer / Axial-DeepLabCFT (ours)
Primary objectiveReduce the cost of long-range attentionExploit cyclic human mobility patterns for anomaly detection
Attention axesImage height × widthWithin-day × across-day
Inductive biasSpatial relationshipsDaily movement episodes and recurring routines
Input representationSpatial image tensorsHTI from dense GPS or sparse stay-points

As acknowledged in Sec. 3.3, CFT builds on axial attention rather than introducing a new attention operator. Adapting existing attention mechanisms to task-specific structures is common practice, as seen in Axial-DeepLab and TimeSformer. Our contribution is the joint design of HTI and CFT, where attention is organized around the temporal structure of human mobility. Unlike a flat Transformer, which attends over all time slots, CFT explicitly models within-day behavior and recurring patterns across days. This inductive bias encourages the model to focus on behaviorally relevant comparisons while also reducing computational cost.

Our ablations in Table 2 show that both attention axes contribute to detection performance, while the calendar-alignment experiment discussed below further demonstrates the importance of aligning these axes with human mobility patterns.

AuthorsOpenReview ↗

General response (Part 1 of 3)

General response

We thank all three reviewers for their careful reading. R-xzE5 and R-WxxM recognize the unification of dense GPS and sparse stay-point trajectories under a single representation as a strength. R-xzE5 and R-Ca8V find the design well motivated by the within-day and across-day patterns of human mobility, and highlight how this inductive bias, captured through intra-day and inter-day attention, improves anomaly detection. R-Ca8V also highlights the gains in temporal localisation, supported by our ablations of both attention axes. All three reviewers acknowledge the computational efficiency of CFT, which is 11–75× faster than a flat Transformer, while R-xzE5 also finds the paper well written and easy to follow.

Four concerns were raised by more than one of you: (1) a held-out test set, (2) a direct comparison with the closest methods and how CFT differs from axial attention, (3) whether the calendar alignment matters beyond factorized attention, and (4) simulated data and synthetic anomalies. We answer them here with new experiments. Every other point is answered in the reply to each review.

1. Held-out test set

R-xzE5 (Q5): "Is there a test split separate from the validation split used for checkpoint selection?"

R-Ca8V (Q1): "Are the results in Table 1 computed on an independent test set?"

R-WxxM (W3): "The paper reports training and validation splits but does not clearly establish a test set."

We thank the reviewers for the suggestion. We now report results on a held-out Test set for every benchmark. To address the concern that model selection and evaluation used the same data, every benchmark now has three disjoint sets of agents: Train, Validation, and a held-out Test set (sizes in Appendix A of the revised paper). We retrained CFT and selected competing baselines from scratch on Train. The Validation set is used only to select the checkpoint and the detection threshold. No tuning of any kind is done on the held-out Test set: it is never seen during training or model selection, and each trained model is evaluated on it once. Due to limited time during the rebuttal period, we focused on comparing our method against the strongest and most relevant baselines and will rerun the remaining methods for the camera-ready version.

The results remain consistent with the paper: the methods reported here have temporal AUC-PR within about 0.04 of their original scores, and CFT remains the best-performing method among those evaluated on the new Test sets, with a margin of at least 0.13 in temporal AUC-PR on benchmarks with temporal anomaly labels. We have also updated the paper to include the training, validation, and test set sizes and reflect the new evaluation protocol.

The Paper column shows the original results, while Test reports performance on the new held-out set. Test mIoU uses the threshold tuned on the Validation set. Dense Tokyo Test scores are computed on 15-minute slots for all methods.

BenchmarkMethodPaper (AUC-PR)Test (AUC-PR)Test mIoU
NumoSim-LA, temporalCFT (ours)0.590.600.73
USTAD0.460.430.65
SegFormer0.330.300.66
NumoSim-LA, agentCFT (ours)0.650.680.76
USTAD0.530.490.65
SegFormer0.620.650.74
PoL, agentCFT (ours)0.710.65 ± 0.030.71
SegFormer0.680.60 ± 0.230.64
Dense Tokyo, temporalCFT (ours)0.840.830.83
Chronos-20.690.700.73
Dense Tokyo, agentCFT (ours)1.001.001.00
Chronos-20.820.860.65
Dense2Sparse, temporalCFT (ours)0.200.170.56
DINOv2+Adapter0.070.040.52
Dense2Sparse, agentCFT (ours)0.780.810.85
DINOv2+Adapter0.670.630.80

What TITAnD contributes that prior work does not

Dear Reviewers,

While answering the experimental questions in the rebuttal, our contribution might have gotten buried. Given experimental questions are now addressed, so we want to state plainly what TITAnD contributes that prior work does not.

Context: this is not typical time-series data. A GPS trajectory is a time series, but not a typical one. Its values are positions on a map, so what matters is not only when, but where the person is, what is there and how they move: time, space, semantics and kinematics together. It is also far longer than what time-series models handle, and it comes in two forms: a dense GPS stream, and sparse events with a start time, an end time and a duration, which are not a regular series at all. Standard time-series models therefore cannot take it as it is. To run Chronos-2 and TimeMixer++ at all, we had to re-engineer the data for them, deviating from their original papers. Those adapted models are the baselines in Table 8, and CFT still beats both at localization. Thus, patching a time-series model is not the answer. In a similar vein, dense and sparse trajectory methods have stayed separate for the same reason: a model built for one cannot be applied to the other.

Our paper does four things for the first time:

1. We are the first to unify dense and sparse trajectory anomaly detection.

  • What's new: HTI puts a dense GPS stream and a sparse stay-point log into the same image, and one architecture reads either. We do this by formulating the task as a vision problem: detection is image classification and localization is semantic segmentation, trained end to end.
  • What it buys: One architecture serves both sparse and dense data. It also opens the door to pre-trained vision models such as DINOv2.

2. We are the first to show how the trajectory image must be built.

  • What's new: We show that when building the trajectory image (HTI), its axes must follow the structure of human mobility, within-day by across-day, to capture this inductive bias. Otherwise performance is poor. This proves HTI is not yet another 1D-to-2D transformation (e.g., TimeMixer++). No prior work has shown that this alignment is what matters for human mobility.
  • What it buys: With the same architecture, performance drops when this structure is broken (Table 9).

3. We are the first to detect anomalies over multi-month dense GPS.

  • What's new: Prior dense methods work on single trips. Long-horizon methods work on stay-points and discard the dense evidence because of its cost. For example, one agent's raw trajectory is 1.55M points at 6 months, far beyond what general time-series models are built for (Chronos-2 supports 8,192 steps). Thus they do not scale. HTI removes this barrier: it compresses the dense GPS into a compact image whose pixels keep its time, space, semantics and kinematics, instead of discarding them.
  • What it buys: Months of dense data become tractable without losing the dense evidence.

4. We are the first to apply axial attention to trajectory anomaly detection.

  • What's new: We are the first to use axial attention for trajectory anomaly detection. Existing methods could not use it at all, because they treat trajectory anomaly detection as a sequence problem; our vision formulation is what allows it. Only the operator is shared with prior work. There, axial attention takes an RGB image, its axes are height and width, and its purpose is to cut computation. In CFT, the input is the HTI, the axes are within-day and across-day, and the purpose is to capture the structure of human mobility.
  • What it buys: The attention acts as a filter. Each query slot attends only to its own day and to the same time on other days, not to the many slots that are irrelevant to it. This inductive bias makes the pattern easier to learn: a flat Transformer that attends to every slot is clearly weaker (Table 1). It is also efficient, 11-75× faster than that flat Transformer.

The table shows where prior work overlaps with TITAnD. Only TITAnD has every ingredient.

Trajectory anomaly detectionDense GPSSparse stay-points2D time gridDay × time-of-day layoutAxis-wise attentionWithin-day × across-day attention
Sparse methods (USTAD, ICAD, LM-TAD)✓✓
Dense trip-level methods (iBAT, GM-VSAE, FOTraj)✓✓
Period-folding time series (TimeMixer++)✓✓
Axial attention (Axial Transformer, Axial-DeepLab)✓
TITAnD (ours)✓✓✓✓✓✓✓

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.