Pith. sign in

REVIEW 3 major objections 6 minor 21 references

Inferring Missing Trajectory Data with Temporal Convolutional Networks

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read A non-causal temporal convolutional network reconstructs contiguous missing trajectory segments from both past and future context, beating linear interpolation on synthetic multi-sinusoid paths.

desk verdict Clean small methods note on non-causal TCNs for synthetic trajectory gaps, but the headline numbers look like they may be train-set metrics and the baseline bar is low. read the letter →

arxiv 2607.25147 v1 pith:WJZJJRDV submitted 2026-07-27 cs.AI

Ilinca Tiriblecea , Gabriel Turinici This is my paper · ORCID
classification cs.AI
keywords trajectoryinpaintingtemporalconvolutionalnetworkssequencemodelingmissingdatareconstructionnon-causaldilationsynthetictrajectoriescompositeloss
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

Real trajectory recordings often contain contiguous gaps from sensor failure, occlusion, or lost packets. This paper argues that the right architecture for filling those gaps is not a forecasting model that only looks backward, but a temporal convolutional network whose dilations are made symmetric so every time step can use both past and future observations. Trained with a loss that weights errors inside the gap, penalizes jumps at the gap edges, and encourages smooth motion, the network recovers most of the structure of held-out segments on a controlled synthetic benchmark of two-dimensional multi-frequency paths. A sympathetic reader cares because trajectory inpainting is a prerequisite for almost every downstream use of motion data, yet it has been studied far less than forecasting, and the paper isolates the architectural and loss choices needed for the reconstruction setting itself.

What carries the argument

A non-causal Temporal Convolutional Network: five residual blocks of dilated 1-D convolutions (dilations 1, 2, 4, 8, 16; kernel size 5; 64 channels) with symmetric zero-padding instead of causal padding, so the receptive field of 125 steps draws equally on past and future context, trained under the composite loss L = L_MSE + 0.5 L_cont + 0.5 L_smooth.

What would settle it

Train the same non-causal TCN and loss on a real pedestrian or wildlife trajectory set with systematic contiguous masks (or natural occlusion gaps), then check whether masked-region R², MSE, and MAE still beat linear interpolation and standard imputation baselines by a comparable margin; collapse of that margin would falsify the claim that the architecture and loss transfer beyond the synthetic regime.

Watch

Extended reading notes

Core claim

On a synthetic set of 1,000 training, 200 validation, and 300 test two-dimensional trajectories built from superimposed sinusoids with random phases and light Gaussian noise, each with one randomly placed contiguous 20 percent mask, a five-block TCN with symmetric (non-causal) dilated convolutions and a composite weighted-MSE plus boundary-continuity plus smoothness loss achieves mean masked-region R² of 0.776, MSE of 0.004, and MAE of 0.047, clearly outperforming linear interpolation.

Load-bearing premise

That strong results on smooth parametric multi-sinusoid trajectories with one fixed-length gap and tiny independent noise are informative about real trajectory dropouts, which have non-stationary dynamics, structured sensor noise, and variable or multiple gaps.

Editorial extensions

If this is right

  • Inpainting-oriented sequence models should drop the causal-padding constraint standard in forecasting TCNs and use symmetric context.
  • Composite losses that separately target gap interiors, gap boundaries, and global smoothness measurably improve reconstruction over plain MSE.
  • Controlled synthetic multi-frequency trajectories are a usable first benchmark for isolating architecture and loss design before real-data evaluation.
  • The same non-causal TCN pattern is a computationally lighter alternative to recurrent or transformer inpainters when the receptive field covers the gap.

Reading between the lines

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

  • If multi-gap and variable-length masking were added, the fixed receptive field of 125 steps would likely force either deeper dilation stacks or hierarchical TCNs—an extension the paper flags but does not test.
  • The large R² standard deviation suggests a natural next metric suite that stratifies error by local curvature or frequency content inside the gap.
  • Boundary-continuity penalties of the form used here could transfer to other gap-filling domains (audio inpainting, missing sensor channels) where edge artifacts dominate perceived quality.
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

3 major / 6 minor

Summary. The manuscript addresses trajectory inpainting — reconstructing contiguous missing segments of 2-D trajectories — using a Temporal Convolutional Network with symmetric (non-causal) dilated convolutions, trained with a composite loss combining weighted MSE over masked vs. observed regions, a boundary-continuity penalty at gap edges, and a first-order smoothness regularizer. On a synthetic benchmark of 1,500 multi-sinusoid trajectories (1,000 train / 200 validation / 300 test) with a single randomly placed 20% contiguous mask, the model is reported to achieve mean masked-region R² = 0.776, MSE = 0.004, MAE = 0.047, outperforming linear interpolation (R² = −0.437). Ablations in Table 3 show modest degradation when the continuity or smoothness terms are removed, and qualitative examples show boundary-consistent completions.

Significance. Trajectory inpainting is genuinely underexplored relative to forecasting, and the problem framing (isolated from detection noise and identity ambiguity via controlled synthetic data) is reasonable and honestly motivated in §2.1. The non-causal TCN adaptation is simple but appropriate for the task, and the paper is unusually transparent: the data generator is fully specified (Eqs. 1–2), the architecture and receptive field (125 steps) are fully specified (Table 1), the loss hyperparameters are stated, and limitations are candidly acknowledged in §3.4. If the headline metrics are confirmed to be test-set numbers, the paper constitutes a modest but sound proof-of-concept; however, it remains a synthetic-only study with a single weak baseline, so its significance for real-world trajectory inpainting is limited, as the authors themselves concede.

major comments (3)
  1. [§3.1, Fig. 1] §3.1 (Figure 1 caption and accompanying text): the headline metrics are described as 'computed per trajectory and then averaged over the full 1,000-sample dataset.' Per §2.1, the 1,000-sample set is the training set; the test set has 300 samples. This directly contradicts the ablation in §3.3, whose Table 3 explicitly states 'test set, 300 samples,' and sits oddly next to Fig. 1 (right), which plots a 'Test Loss' curve, showing a held-out set was monitored. Either (a) the headline comparison against linear interpolation is a train-set measurement — in which case R² = 0.776 for a ~200k-parameter network trained for 50,000 gradient steps on 1,000 trajectories of a fixed two-component parametric family substantially overstates generalization — or (b) the metrics were computed on the test set and the text/caption is mislabeled. This ambiguity is load-bearing for the paper's central quantitat
  2. [§3.1, §4] The only quantitative baseline is linear interpolation. This is a straw-man baseline for trajectories explicitly constructed (Eqs. 1–2) to have curvature 'that cannot be recovered by linear interpolation'; the negative baseline R² = −0.437 is essentially guaranteed by the benchmark design and therefore carries little information about the method's value. The paper itself cites V-spline [20] and the TrajImpute benchmark [5] in §1.1/§2.1. At minimum, a cubic smoothing spline (or V-spline) baseline and one learned baseline (e.g., a bidirectional LSTM or the causal version of the same TCN, to isolate the contribution of symmetric padding) are needed for the claim that 'TCNs are a viable, computationally efficient alternative to recurrent and attention-based models' (§4) to be supported. Given the fully synthetic setup, these additions are inexpensive and within scope.
  3. [§3.3, Table 3] §3.3 and §4 are internally inconsistent about the ablation. The Conclusion lists 'a systematic ablation of the loss components to quantify the individual contribution of the continuity and smoothness terms' as future work, yet Table 3 already presents exactly that ablation. More substantively, Table 3 shows degradations of only ~0.02–0.03 in mean R² with standard deviations (~0.8–0.9) an order of magnitude larger than the effect, and the text itself acknowledges the conclusion 'depends... on the specific run of the algorithm.' With a single stochastic training run per configuration, the ablation does not support the claim that the continuity and smoothness terms matter. Either multiple seeds with confidence intervals, or a softening of the claim, is required.
minor comments (6)
  1. [§2.4] The choice λ_cont = λ_smooth = 0.5 (and α = 0.1 in Eq. 5) is asserted without justification or sensitivity analysis; a brief sweep or a statement of how these were selected would strengthen §2.4.
  2. [§3.1] Per-trajectory R² has std 0.777 around a mean of 0.776 and the linear-interpolation std is 5.639; the text mentions median scores but does not report them numerically. Reporting medians (and possibly IQR) in the main table would be more informative for such skewed distributions.
  3. [§3.1, Fig. 1] Metrics in §3.1 are presented inside a figure environment (Fig. 1) rather than a numbered table, making them awkward to reference; converting to a table (as done for Tables 2–3) would improve clarity.
  4. [§3.1, Eq. (10)] The R² definition in Eq. (10) uses the masked-segment mean p̄ as the reference; since segment length and position vary per trajectory, the per-trajectory R² values are not on a common scale, which partially explains the large variance. Worth one sentence of discussion.
  5. Several typographical/spacing artifacts appear throughout (e.g., 'adoptsymmetric', 'remainthemostwidelyused', 'Thiscontrolledsettingenablespreciseanalysis', 'F unction', 'NN architecture' → presumably 'network architecture' in §2.3); a proofreading pass is needed.
  6. [§2.1] The mask is fixed at 40 steps (20%) throughout; since the receptive field is 125 steps, a sentence noting how performance scales with gap length (or why 20% was chosen) would help readers assess the operating regime.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: supervised TCN inpainting on independent synthetic trajectories; metrics are held-out reconstruction error, not inputs renamed as predictions.

full rationale

This is a standard empirical sequence-modeling paper. Synthetic trajectories are generated from fixed multi-sinusoid formulas (Eqs. 1–2) with random phases and noise, independent of the network. The TCN (symmetric dilated convolutions, Table 1) and composite loss (Eqs. 4–7) are design choices with fixed hyperparameters (α=0.1, λ_cont=λ_smooth=0.5), not quantities fitted then re-reported as predictions. Evaluation metrics (MSE/MAE/R² on masked indices, Eqs. 8–10) measure reconstruction of held-out coordinates never supplied as targets at test time; linear interpolation is an external baseline. Citations (Bai et al. TCN, Social-LSTM, TrajImpute, etc.) are external literature; there is no self-citation load-bearing uniqueness claim, no ansatz smuggled from the authors’ prior work, and no renaming of a known closed-form result. Any train/test reporting inconsistency noted by the skeptic is an evaluation-integrity concern, not circularity of derivation. Derivation chain is self-contained and non-circular.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

Load-bearing content is almost entirely standard ML practice plus an explicit synthetic process. The central empirical claim rests on the assumption that the multi-sinusoid generator and single-gap 20% mask are a fair testbed, on hand-chosen loss weights, and on the architectural hyper-parameters (depth, dilations, channels, kernel). No new physical entities are postulated.

free parameters (5)
  • lambda_cont, lambda_smooth = 0.5 each
    Continuity and smoothness loss weights fixed at 0.5 without sweep reported; ablation only tests on/off.
  • alpha (observed-region MSE weight) = 0.1
    Down-weights reconstruction loss on known points; chosen as 0.1, not cross-validated in the text.
  • TCN width/depth/kernel/dilations = 64 ch, k=5, d=2^{ℓ-1}
    5 blocks, 64 channels, k=5, dilations 1,2,4,8,16 chosen by authors; receptive field 125 is a design choice that must cover the 40-step gaps.
  • noise sigma and mask fraction = σ=0.02, mask=20%
    Generator noise σ=0.02 and 20% single contiguous mask define the entire benchmark difficulty.
  • Adam learning rate and epoch count = 3e-3, 50 epochs, bs=1
    lr=3e-3, 50 epochs, batch size 1; training schedule is part of the reported performance.
assumptions (4)
  • domain assumption Stacked dilated 1D convolutions with residual connections and layer norm are a stable sequence model (Bai et al. TCN framework).
    Architecture section builds directly on the standard TCN block design without re-deriving its approximation properties.
  • domain assumption Symmetric (non-causal) padding is valid and desirable for inpainting because both past and future context are observed at fill-in time.
    Stated in Introduction and §2.3 as the key design relaxation versus forecasting TCNs.
  • ad hoc to paper Synthetic multi-sinusoid trajectories with i.i.d. Gaussian noise and one random contiguous gap adequately isolate the inpainting problem for method comparison.
    §2.1 Motivation and Technical details; all reported metrics depend on this generator. Authors note in §3.4 that real data remains open.
  • ad hoc to paper Linear interpolation is a sufficient primary baseline to demonstrate value of the learned model.
    §3.1 Baseline; no spline, Kalman, RNN, or transformer imputation baseline is reported on the same split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inferring Missing Trajectory Data with Temporal Convolutional Networks." pith.science (2026). https://pith.science/paper/WJZJJRDV

@misc{pith2026260725147,
  author       = {Pith},
  title        = {Pith review of: Inferring Missing Trajectory Data with Temporal Convolutional Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJZJJRDV}},
  note         = {Machine review of arXiv:2607.25147}
}
abstract

Trajectory data collected in real-world settings is frequently incomplete due to sensor failure, communication loss, or occlusion. We address the task of \emph{trajectory inpainting}: reconstructing contiguous missing segments from observed context. We propose a Temporal Convolutional Network (TCN) with symmetric dilation that relaxes the standard causality constraint, allowing each time step to draw on both past and future observations, a property that is essential for inpainting, but absent from forecasting-oriented architectures. The model is trained with a composite loss that combines weighted mean squared error, boundary--continuity penalties, and a smoothness regularizer. Trained on a synthetic dataset of $1,000$ (train), $200$ (validation), and $300$ (test) two-dimensional trajectories with randomly placed 20% masked segments, the model achieves good R$^{2}$, MSE and MAE metrics.

Figures

Figures reproduced from arXiv: 2607.25147 by the authors.

Figure 1
Figure 1. Left top: Quantitative evaluation on masked regions (trained on 1,000 syn￾thetic trajectories, 20% mask fraction) for a linear interpolation baseline. Left bottom: Quantitative evaluation on masked regions (trained on 1,000 synthetic trajectories, 20% mask fraction). Right: Loss evolution on the train and test datasets. Dataset Metric Mean ± Std Smooth MSE 0.0011 ± 0.0005 MAE 0.025 ± 0.006 R 2 0.984 ± 0.006 Dataset … view at source ↗
Figure 2
Figure 2. True versus predicted trajectories. The predicted trajectories are constructed using true values for the known part and predicted values for the masked part. Metric Mean ± Std MSE 0.005 ± 0.018 MAE 0.048 ± 0.031 R 2 0.743 ± 0.831 Metric Mean ± Std MSE 0.005 ± 0.013 MAE 0.049 ± 0.029 R 2 0.752 ± 0.908 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 2 canonical work pages

  1. [20]

    In: Proceedings of the 1st ACM SIGSPATIAL in- ternational workshop on AI-driven spatio-temporal data analysis for wildlife con- servation

    Wan, Z., Dodge, S.: A generative trajectory interpolation method for imputing gaps in wildlife movement data. In: Proceedings of the 1st ACM SIGSPATIAL in- ternational workshop on AI-driven spatio-temporal data analysis for wildlife con- servation. pp. 1–8 (2023)

    2023

  2. [5]

    In: Proceedings of the 38th International Confer- ence on Neural Information Processing Systems (NeurIPS)

    Chib, P.S., Singh, P.: Pedestrian trajectory prediction with missing data: Datasets, imputation, and benchmarking. In: Proceedings of the 38th International Confer- ence on Neural Information Processing Systems (NeurIPS). pp. 124530–124546. NIPS ’24, ACM (2024)

    2024

  3. [1]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social lstm: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961–971 (2016)

    2016

  4. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Bae, I., Park, Y.J., Jeon, H.G.: Singulartrajectory: Universal trajectory predictor using diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17890–17901 (2024)

    2024

  5. [3]

    arXiv preprint arXiv:1803.01271 (2018), https://arxiv.org/abs/1803.01271

    Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271 (2018), https://arxiv.org/abs/1803.01271

    arXiv 2018

  6. [4]

    Sensors21(9), 3215 (2021)

    Cao, Z., Bryant, D., Molteno, T.C., Fox, C., Parry, M.: V-spline: An adaptive smoothing spline for trajectory reconstruction. Sensors21(9), 3215 (2021). https: //doi.org/10.3390/s21093215

    work page doi:10.3390/s21093215 2021

  7. [6]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social gan: Socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2255–2264 (2018)

    2018

  8. [7]

    Nature229, 381–383 (1971)

    Henderson, L.F.: The statistics of crowd fluids. Nature229, 381–383 (1971)

    1971

Show all 21 references
  1. [8]

    Neural Computation 9(8), 1735–1780 (1997)

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735–1780 (1997)

    1997

  2. [9]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  3. [10]

    In: Proceedings of the European Confer- ence on Computer Vision (ECCV) Workshops

    Lea, C., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks: A unified approach to action segmentation. In: Proceedings of the European Confer- ence on Computer Vision (ECCV) Workshops. Lecture Notes in Computer Science, vol. 9915, pp. 47–54. Springer, Cham (2...

    2016 arXiv

  4. [11]

    arXiv preprint arXiv:1504.01942 (2015) Inferring Missing Trajectory Data with Temporal Convolutional Networks 11

    Leal-Taixé, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942 (2015) Inferring Missing Trajectory Data with Temporal Convolutional Networks 11

    2015 arXiv

  5. [12]

    In: Computer graphics forum

    Lerner, A., Chrysanthou, Y., Lischinski, D.: Crowds by example. In: Computer graphics forum. vol. 26, pp. 655–664. Wiley Online Library (2007)

    2007

  6. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR)

    Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-stgcnn: A social spatio- temporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 14424–14432 (2020)

    2020

  7. [14]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    Pellegrini,S.,Ess,A.,VanGool,L.:You’llneverwalkalone:Modelingsocialbehav- ior for multi-target tracking. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 261–268 (2009)

    2009

  8. [15]

    International Journal of Robotics Research39(8), 895–935 (2020)

    Rudenko, A., Palmieri, L., Herman, M., Kitani, K., Gavrila, D.M., Arras, K.O.: Human motion trajectory prediction: A survey. International Journal of Robotics Research39(8), 895–935 (2020)

    2020

  9. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shi, L., Wang, L., Long, C., Zhou, S., Zhou, M., Niu, Z., Hua, G.: Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8994– 9003 (2021)

    2021

  10. [17]

    In: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Sun, G., Wang, S., Zhu, L., Liu, M., Ma, J.: Gdts: Goal-guided diffusion model with tree sampling for multi-modal pedestrian trajectory prediction. In: 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 14595– 14602. IEEE (2025)

    2025

  11. [18]

    Journal of Ambient Intelligence and Humanized Computing17(2), 491– 523 (Mar 2026)

    Taha, M., Zaky, A., Yamaguchi, H., Fares, A.: Pedestrian trajectory and in- tention prediction: a comprehensive review of models, datasets, and chal- lenges. Journal of Ambient Intelligence and Humanized Computing17(2), 491– 523 (Mar 2026). https://doi.org/10.1007/s12652-026-0...

    2026 doi

  12. [19]

    In: Advances in Neural Information Processing Systems

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30 (2017)

    2017

  13. [21]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

    Yuan, Y., Weng, X., Ou, Y., Kitani, K.: Agentformer: Agent-aware transform- ers for socio-temporal multi-agent forecasting. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9793–9803 (2021). https://doi.org/ 10.1109/ICCV48922.2021.00967

    2021

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.